WEBVTT Kind: captions Language: en 00:00:00.250 --> 00:00:01.680 Good morning. 00:00:01.680 --> 00:00:04.460 Welcome to today’s Earthquake Science seminar. 00:00:04.460 --> 00:00:07.680 And kind of welcome back to our regularly scheduled science seminars. 00:00:07.680 --> 00:00:10.020 A couple of quick announcements before we start today. 00:00:10.020 --> 00:00:14.720 Tomorrow afternoon, starting at 2:15, but we’ll have coffee starting at 2:00, 00:00:14.720 --> 00:00:19.140 in California, which is across the way, James Malley and Robert Pekelnicky 00:00:19.140 --> 00:00:22.440 from Degenkolb Engineers will talk about seismic performance 00:00:22.440 --> 00:00:25.520 of tall buildings in San Francisco. And that will kind of wrap up our 00:00:25.520 --> 00:00:33.460 month-long series on seismic design and tall-building response in the Bay Area. 00:00:33.460 --> 00:00:37.400 And the – sorry. Next week, we will have 00:00:37.400 --> 00:00:42.480 Sara Dougherty here in Rambo auditorium, but at 1:30 in the afternoon. 00:00:43.220 --> 00:00:45.300 We’ve been moved back because of the town hall that will 00:00:45.309 --> 00:00:48.980 be happening at 11:00 next week. But today, we’re lucky to have 00:00:48.980 --> 00:00:53.040 Artie Rodgers over from Lawrence Livermore, and Brad will introduce him. 00:00:55.020 --> 00:01:00.000 - Artie received his B.S. in physics from Northeastern University, 00:01:00.000 --> 00:01:03.220 and then a Ph.D. from University of Colorado in Boulder. 00:01:03.220 --> 00:01:07.900 After that, he was a postdoc at New Mexico State and UC-Santa Cruz. 00:01:07.900 --> 00:01:11.500 He joined the seismology group at Lawrence Livermore National Lab in 00:01:11.500 --> 00:01:18.240 1997 as a postdoc and was a group seismology leader from 2006 to 2010. 00:01:18.240 --> 00:01:20.740 And that’s when we started to collaborate with Artie 00:01:20.740 --> 00:01:23.210 when we were doing the ground motions for the 00:01:23.210 --> 00:01:27.660 1906 centennial as well as our Hayward scenarios. 00:01:27.660 --> 00:01:32.450 Artie is an expert in nuclear explosion monitoring, crustal and lithosphere 00:01:32.450 --> 00:01:37.460 structure, regional seismic wave propagation, and event identification. 00:01:37.460 --> 00:01:41.520 And in 2010, he was a Fulbright Scholar at Grenoble in France. 00:01:41.520 --> 00:01:45.680 And in addition to being a technical staff member at Lawrence Livermore, 00:01:45.680 --> 00:01:48.420 he is a visiting scientist at UC-Berkeley, 00:01:48.420 --> 00:01:51.520 an affiliate of Lawrence Berkeley National Labs. 00:01:51.520 --> 00:01:55.140 And today, Artie is going to talk about high-frequency simulations 00:01:55.149 --> 00:01:57.530 of scenario earthquakes on the Hayward Fault. 00:01:57.530 --> 00:01:59.760 - All right. Thanks, Brad. 00:02:00.820 --> 00:02:05.180 Thanks for having me here. This – originally we had scheduled 00:02:05.180 --> 00:02:10.310 to visit in June, and then we had an opportunity to run some calculations on 00:02:10.310 --> 00:02:18.200 the Cori Cluster at Lawrence Berkeley Lab, and that collided with the date. 00:02:18.200 --> 00:02:21.520 But fortunately, you know, that all went well, and thanks for rescheduling. 00:02:21.520 --> 00:02:25.500 I’ll show results from that calculation and then from another calculation 00:02:25.500 --> 00:02:29.880 that we’ve done more recently at even higher resolution. 00:02:29.880 --> 00:02:31.970 So I’m going to focus on high-performance computing 00:02:31.970 --> 00:02:36.440 simulations of ground motions for a Hayward Fault scenario. 00:02:36.440 --> 00:02:42.040 And I’ve worked on some other topics that are relevant, but I won’t be able to 00:02:42.040 --> 00:02:47.230 talk to today, but some of those themes will come up. 00:02:47.230 --> 00:02:50.560 This is work that’s supported by the Department of Energy’s 00:02:50.560 --> 00:02:53.430 Exascale Computing Project, which is trying to develop hardware 00:02:53.430 --> 00:02:56.620 and software for the next generation of supercomputers – 00:02:56.620 --> 00:03:01.640 exascale, or a billion billion flops. And the current fastest machines 00:03:01.650 --> 00:03:07.620 are sort of 10 to 100 petaflops. So this is a big DOE effort to 00:03:07.620 --> 00:03:12.560 be able to take advantage of the next generation of supercomputers. 00:03:12.560 --> 00:03:16.069 So sort of the one-slide summary of this is, I’ll talk about these 00:03:16.069 --> 00:03:19.260 high-performance computing ground motion simulations 00:03:19.260 --> 00:03:22.810 of a magnitude 7 earthquake on the Hayward Fault, which is shown – 00:03:22.810 --> 00:03:28.450 on the PGV map, is shown here as the background and selected seismograms. 00:03:28.450 --> 00:03:32.830 These are – use the SW4 finite different code. 00:03:32.830 --> 00:03:36.740 It’s fully three-dimensional physics-based wave propagation 00:03:36.740 --> 00:03:41.120 using the USGS 3D geologic, slash, seismic model. 00:03:41.120 --> 00:03:45.690 We include surface topography and anelastic attenuation. 00:03:45.690 --> 00:03:50.640 Our goal is to do full deterministic 3D calculations that are broadband – 00:03:50.640 --> 00:03:54.220 so without having to use any hybrid methods to add the high frequencies – 00:03:54.220 --> 00:03:58.040 no stochastic method, no 1D Green's functions – 00:03:58.040 --> 00:04:02.900 just purely physics-based calculation. And we’re aiming to get as 00:04:02.900 --> 00:04:07.849 high frequency as we can, and we’ve gotten to 2-1/2, and now 6.9 hertz. 00:04:07.849 --> 00:04:11.620 And this requires very fine discretization to be able to resolve the very short 00:04:11.620 --> 00:04:16.920 wavelength waves on the order of, you know, 12 to 9 meters. 00:04:16.920 --> 00:04:20.769 And in order to run over this large domain with such fine resolution, 00:04:20.769 --> 00:04:24.080 we need to run on large supercomputer systems. 00:04:24.880 --> 00:04:28.100 We’ve analyzed the results that come out of these calculations, 00:04:28.100 --> 00:04:32.050 and they agree well with ground motion models, and I’ll show some of that. 00:04:32.050 --> 00:04:35.910 And there’s work to be done to evaluate the path and site effects. 00:04:35.910 --> 00:04:41.169 There’s an asymmetry in the PGV across the Hayward Fault that often shows up, 00:04:41.169 --> 00:04:44.780 going back to the calculations with did with Brad in 2010. 00:04:44.780 --> 00:04:48.120 And we’ll maybe touch on that a bit. 00:04:48.139 --> 00:04:53.500 The work we’re doing is supporting engineering analysis of structures. 00:04:53.500 --> 00:04:59.550 So the project we’re on has the component of both 00:04:59.550 --> 00:05:03.190 building response and earthquake soil structure interaction. 00:05:03.190 --> 00:05:07.310 So we’re generating motions here, and then providing that to our 00:05:07.310 --> 00:05:09.360 engineering colleagues at Lawrence Berkeley Lab 00:05:09.360 --> 00:05:12.320 and UC-Davis to perform these engineering analyses. 00:05:12.320 --> 00:05:18.350 So I won’t talk more about that, but it’s a nice part of this project that, 00:05:18.350 --> 00:05:21.440 you know, we’re in service to the engineering here, and they’re 00:05:21.440 --> 00:05:26.060 keeping us honest in terms of the ground motions we’re calculating. 00:05:26.060 --> 00:05:28.830 This is a team effort. You know, I’ve been working for 00:05:28.830 --> 00:05:31.710 the last – almost – more than 12 years with Anders Petersson and 00:05:31.710 --> 00:05:35.900 Bjorn Sjogreen at Livermore to – they’re the applied mathematicians 00:05:35.900 --> 00:05:39.830 that have been developing the SW4 mathematical methodology 00:05:39.830 --> 00:05:42.290 as well as the computer code. 00:05:42.290 --> 00:05:47.900 Recently, Bjorn and Ramesh at Livermore have ported the 00:05:47.900 --> 00:05:51.759 SW4 code from a CPU-based system to a GPU system. 00:05:51.759 --> 00:05:56.169 That’s effort that’s been ongoing for three years, and that’s borne fruit. 00:05:56.169 --> 00:06:00.860 And I’ll show results with that from the Sierra supercomputer at Livermore. 00:06:00.860 --> 00:06:04.060 I’ve been working with Arben Pitarka since he joined Livermore, 00:06:04.060 --> 00:06:06.810 and he’s my office mate and somebody I talk to every day 00:06:06.810 --> 00:06:10.330 about these, and he’s informed a lot of what we’re doing. 00:06:10.330 --> 00:06:13.170 And Rob Graves in Pasadena has provided 00:06:13.170 --> 00:06:15.680 his rupture generator, which we use. 00:06:15.680 --> 00:06:19.520 Dave McCallen is the PI of this project, and he’s a structural engineer. 00:06:19.530 --> 00:06:26.949 And he and his co-workers are working to analyze the structural analysis 00:06:26.949 --> 00:06:30.590 with the motions that we’ve calculated. And Dave has created a number of 00:06:30.590 --> 00:06:34.630 advisory boards and working groups, and a lot of those people 00:06:34.630 --> 00:06:37.319 have helped out a lot. But Norm Abrahamson in particular 00:06:37.319 --> 00:06:43.139 has helped inform us on acceptance criteria for ground motions and 00:06:43.140 --> 00:06:49.680 how to incorporate simulation results into next-generation hazard thinking. 00:06:49.680 --> 00:06:52.090 And none of what I’ve talked about would be possible without 00:06:52.090 --> 00:06:56.100 the supercomputing centers, which are, you know, run by hundreds 00:06:56.100 --> 00:06:59.090 of people at the NERSC facility at Lawrence Berkeley Lab 00:06:59.090 --> 00:07:01.720 and Livermore Computing at Lawrence Livermore Lab. 00:07:02.160 --> 00:07:06.199 Okay, so we all know – you know, you guys know this better than I do. 00:07:06.200 --> 00:07:10.760 But the Hayward Fault presents – poses great hazard to the East Bay 00:07:10.760 --> 00:07:15.400 and exposes, you know, what, 2-1/2 billion – million people 00:07:15.400 --> 00:07:22.040 to strong ground shaking, cutting through the densely populated East Bay. 00:07:22.040 --> 00:07:26.280 And in 1868, there was, you know, the last major earthquake 00:07:26.280 --> 00:07:29.800 on the Hayward Fault, but only 3,000 people lived there. 00:07:29.800 --> 00:07:32.910 The motions were strong enough to cause structural damage. 00:07:32.910 --> 00:07:36.150 And if such an event were to happen today, 00:07:36.150 --> 00:07:40.780 it would be quite devastating, economically, socially, et cetera. 00:07:41.640 --> 00:07:44.259 And there’s overwhelming evidence that there have been many of 00:07:44.259 --> 00:07:48.240 these events through the paleoseismic record with about 00:07:48.240 --> 00:07:53.289 150 or – you know, 140- to 160-year return periods. 00:07:53.289 --> 00:07:57.280 So here we are entering the – you know, the anniversary – the 150th anniversary. 00:07:57.280 --> 00:08:01.990 So it’s a real timely – you know, it’s a very important time 00:08:01.990 --> 00:08:06.660 to be considering this event and potential consequences. 00:08:07.180 --> 00:08:10.240 So why do we simulate, you know, near-field ground motions? 00:08:10.250 --> 00:08:16.030 I mean, it’s important we do that because there’s very few recordings 00:08:16.030 --> 00:08:19.000 of very large earthquakes at short distance. 00:08:19.000 --> 00:08:24.040 And motions that you – that are measured there are highly variable. 00:08:24.040 --> 00:08:28.120 Near-fault motions are highly variable because of variations in the slip 00:08:28.129 --> 00:08:30.009 and directivity, rise time, et cetera. 00:08:30.009 --> 00:08:35.380 These are a number of simulated records with the same Boore-Joyner distance. 00:08:35.380 --> 00:08:37.340 And you can see there’s a lot of variability. 00:08:37.349 --> 00:08:40.550 You know, and you have the displacement step and the velocity pulse, 00:08:40.550 --> 00:08:44.950 and that plays into the – how it impacts structures. 00:08:44.950 --> 00:08:49.160 And, of course, coupling into sedimentary structures is important 00:08:49.160 --> 00:08:52.410 as well, so – and that’s all stuff that we can do in the safety 00:08:52.410 --> 00:08:54.760 of our computers without having to, you know, 00:08:54.760 --> 00:08:59.310 wait around for the ground motions to be recorded. 00:08:59.310 --> 00:09:02.690 And then sometimes you have to be right on the structures, 00:09:02.690 --> 00:09:04.450 or right on the faults, right? 00:09:04.450 --> 00:09:09.350 We need to have – we need to have structures near faults, whether it’s critical 00:09:09.350 --> 00:09:16.210 infrastructure such as power plants or, you know, hospitals, et cetera. 00:09:16.210 --> 00:09:18.540 But also, all of our transportation infrastructure 00:09:18.540 --> 00:09:23.050 and lifelines cross these faults. So that’s important as well. 00:09:23.050 --> 00:09:25.690 So sometimes we just need to, you know, understand what those 00:09:25.690 --> 00:09:29.390 motions are and to be able to mitigate the effects. 00:09:29.390 --> 00:09:33.220 So a numerical simulation is appealing. It allows us to do the experiments 00:09:33.220 --> 00:09:37.550 we can’t wait for or perform in nature. So we solve the elastodynamic equations 00:09:37.550 --> 00:09:42.330 of motion and the sort of physics-based numerical simulation. 00:09:42.330 --> 00:09:46.149 Solving F equals m-a for a solid with attenuation, and these methods 00:09:46.149 --> 00:09:49.590 are efficiently parallelized so we can run them at high resolution. 00:09:49.590 --> 00:09:53.980 SW4 includes fully 3D geologic heterogeneity – variations in the 00:09:53.980 --> 00:09:58.570 P- and S-wave velocities, densities, and attenuation factors and includes surface 00:09:58.570 --> 00:10:02.280 topography, which becomes important as we go to higher frequencies. 00:10:02.280 --> 00:10:05.180 Anelastic attenuation – it has mesh refinement that allows the mesh 00:10:05.180 --> 00:10:08.370 to get coarser with depth, which improves the efficiency and 00:10:08.370 --> 00:10:13.140 absorbing boundary conditions, so waves don’t bounce back into the media. 00:10:13.870 --> 00:10:17.200 And this results in the time histories of the motion. 00:10:17.200 --> 00:10:21.620 How high of frequency depends on what computing resources you have. 00:10:21.620 --> 00:10:24.110 But the motions can be used to look at ground failure, 00:10:24.110 --> 00:10:27.000 liquefaction, structural analysis, et cetera. 00:10:27.000 --> 00:10:31.510 So – but, you know, simulations are challenging, though. 00:10:31.510 --> 00:10:33.839 And there are three factors that impact this. 00:10:33.839 --> 00:10:36.610 One is mathematical. We really need to have an accurate, 00:10:36.610 --> 00:10:41.000 stable, and efficient method. And our solution here is to use 00:10:41.000 --> 00:10:45.110 a provably stable method – this summation-by-parts method – 00:10:45.110 --> 00:10:49.170 and use higher-order differencing operators and mesh refinement to 00:10:49.170 --> 00:10:59.550 improve the efficiency and to verify solutions against analytical test cases, 00:10:59.550 --> 00:11:04.779 but also the method of manufactured solutions, which allows verification 00:11:04.779 --> 00:11:11.100 of a problem in a 3D – it verifies the solution in a 3D medium. 00:11:11.100 --> 00:11:13.860 Then there are computational challenges. In order to compute – you know, 00:11:13.860 --> 00:11:15.130 for – we’re looking at large earthquakes. 00:11:15.130 --> 00:11:19.230 You know, a magnitude 7 earthquake can be 70 kilometers long. 00:11:19.230 --> 00:11:21.860 We need large domains. We want to get to high-frequency. 00:11:21.860 --> 00:11:26.330 We need fine discretization. So here, we really have to use 00:11:26.330 --> 00:11:29.589 more powerful computers, but then also optimize the code 00:11:29.589 --> 00:11:35.339 to be efficient on that hardware so we make the best use of that resource. 00:11:35.340 --> 00:11:38.520 And the most challenging are really the physical challenges. 00:11:38.520 --> 00:11:42.840 You know, and that breaks into both the earthquake and the Earth model 00:11:42.840 --> 00:11:47.340 And, you know, we need to generate rupture models that are realistic. 00:11:47.340 --> 00:11:55.760 And our solution here is to use these physics-based models that are based on 00:11:55.760 --> 00:12:03.240 analysis of – empirical analysis of finite fault solutions, but also that 00:12:03.240 --> 00:12:07.430 are informed by rupture dynamics to look at the scaling between slip 00:12:07.430 --> 00:12:11.760 and rise time and things like that. And we need to validate – you know, 00:12:11.760 --> 00:12:17.889 use methods for source generation that are validated in terms of 00:12:17.889 --> 00:12:20.980 the empirical data that’s encoded in the ground motion 00:12:20.980 --> 00:12:23.640 prediction equations or ground motion models. 00:12:23.640 --> 00:12:27.570 And then a bigger challenge is really the accuracy of the 3D subsurface structure. 00:12:27.570 --> 00:12:29.209 You know, we’re using billions of grid points. 00:12:29.209 --> 00:12:33.200 We don’t know billions of things about the Earth at that scale. 00:12:33.200 --> 00:12:37.040 We don’t know the, you know, P wave velocity or the S wave velocity on the 00:12:37.040 --> 00:12:42.380 9-meter grid resolution that we’re using. So that’s a – that’s a big challenge. 00:12:42.380 --> 00:12:45.750 And ultimately, we’d like to improve the existing 3D model through data 00:12:45.750 --> 00:12:51.080 inversion or by using stochastic models. And, you know, very interested to hear – 00:12:51.080 --> 00:12:54.930 well, we’re making heavy use of the Bay Area model that’s been 00:12:54.930 --> 00:12:59.690 developed here in this office. And we’re interested in how, 00:12:59.690 --> 00:13:04.820 you know, that is progressing and how we can, you know, 00:13:04.820 --> 00:13:07.300 make use of the best possible model. 00:13:07.300 --> 00:13:11.540 But it’s possible to validate with observed data, and we have magnitude – 00:13:11.540 --> 00:13:15.800 you know, moderate earthquakes – 4s and 3s and 5s and such. 00:13:15.800 --> 00:13:18.769 And we’ve done some work on that. I won’t talk about that today, but that’s 00:13:18.769 --> 00:13:23.120 something we’ve been working on to try to test the model as well. 00:13:23.120 --> 00:13:28.290 But all of this, especially these physical challenges, really drive the need to have 00:13:28.290 --> 00:13:31.590 an accurate, efficient method so we can run lots of simulations, 00:13:31.590 --> 00:13:38.320 whether it’s to sample, you know, a parameter suite of rupture models 00:13:38.320 --> 00:13:44.310 for a given scenario and to look at ensemble behavior or to run an inverse, 00:13:44.310 --> 00:13:47.209 you know, adjoint wave – full wave inversion problem. 00:13:47.209 --> 00:13:50.480 We need an accurate, efficient method there to do some 00:13:50.480 --> 00:13:52.960 sampling of stochastic heterogeneity. 00:13:52.960 --> 00:13:56.930 So that drives the need for the more efficient method. 00:13:56.930 --> 00:14:03.570 So our goal is to get purely deterministic broadband motions, you know, 00:14:03.570 --> 00:14:05.959 by running these high-performance computing simulations. 00:14:05.959 --> 00:14:08.149 And the ingredients that go in that are the ruptures, 00:14:08.149 --> 00:14:10.610 and we’re using the Graves and Pitarka method. 00:14:10.610 --> 00:14:15.140 For the 3D Earth model, we’re using the Bay Area model developed here. 00:14:15.140 --> 00:14:21.100 We’re using SW4, which is the 3D finite difference code developed at Livermore, 00:14:21.100 --> 00:14:25.200 that is being now extended as part of this Exascale Computing Project. 00:14:25.200 --> 00:14:29.620 It’s available through CIG. And we’re running on high-performance 00:14:29.620 --> 00:14:37.070 machines at the DOE labs and post- processing that data to make sense of it. 00:14:37.070 --> 00:14:41.750 So SW4 – but, right, so our overall goal here is working towards being able to 00:14:41.750 --> 00:14:47.220 computer higher frequencies in a realistic Earth model with shorter run times. 00:14:47.220 --> 00:14:51.209 So SW4 is a fourth-order finite difference code for 00:14:51.209 --> 00:14:55.310 seismic wave propagation. Uses a summation-by-parts method, 00:14:55.310 --> 00:14:59.580 which is a provably stable and accurate energy-conserving method. 00:14:59.580 --> 00:15:02.959 It’s a node-centered second-order wave equation. 00:15:02.959 --> 00:15:06.589 It solves the displacement formulation, and it has super-grid boundary 00:15:06.589 --> 00:15:09.329 conditions, which is efficient at preventing waves from 00:15:09.329 --> 00:15:13.520 bouncing back from the – from the edges of the domain. 00:15:13.520 --> 00:15:16.711 A real key feature is this mesh refinement. 00:15:16.711 --> 00:15:22.209 So it uses a curvilinear mesh that honors the surface topography here. 00:15:22.209 --> 00:15:27.899 And it stretches in the vertical direction. And then connects up with a fixed 00:15:27.899 --> 00:15:32.149 grid spacing cartesian domain here in the blue, and then we can double 00:15:32.149 --> 00:15:37.690 the mesh size, or the grid spacing, at user-specified intervals here. 00:15:37.690 --> 00:15:43.019 And, as the wave speeds increase with depth, then we can have larger 00:15:43.019 --> 00:15:49.519 grid spacings and have more efficient – you know, and if you were to use 00:15:49.519 --> 00:15:53.680 a small – very small grid spacing, your time step would go to very short, 00:15:53.680 --> 00:15:56.340 and then you’d have to run the calculation much, much longer. 00:15:56.340 --> 00:15:58.740 So that’s really key. 00:15:58.740 --> 00:16:03.080 What ends up happening is, this curvilinear mesh that honors 00:16:03.080 --> 00:16:06.360 the topography ends up having most of the grid points. 00:16:06.360 --> 00:16:09.290 You know, like sometimes as much as 90 – you know, 00:16:09.290 --> 00:16:12.209 90% of the grid points are in the curvilinear mesh. 00:16:12.209 --> 00:16:15.431 And Anders and Bjorn are working on methods to now have mesh refinement 00:16:15.431 --> 00:16:17.720 within that curvilinear mesh, especially in the Bay Area. 00:16:17.720 --> 00:16:23.209 You go from Mount Diablo at 1,200 meters to the offshore 00:16:23.209 --> 00:16:27.850 Golden Gate, you know, below sea level. And that means – that variation 00:16:27.850 --> 00:16:31.910 in topography means we need to make this boundary where the 00:16:31.910 --> 00:16:37.040 curvilinear mesh goes into the cartesian mesh at a deeper level. 00:16:37.040 --> 00:16:42.180 And then there’s been work under the ECP to run on machines 00:16:42.180 --> 00:16:46.910 that have many cores per node. So the Cori machine at Lawrence 00:16:46.910 --> 00:16:50.600 Berkeley Lab has 68 cores per node. And there’s been a – there was a lot of 00:16:50.600 --> 00:16:59.079 effort recently to be able to run on that architecture efficiently. 00:16:59.079 --> 00:17:01.700 So here are some waveforms from various calculations. 00:17:01.700 --> 00:17:07.770 From the bottom, they’re low-resolved at 0.3 hertz, and then 0.6, and 1.25 00:17:07.770 --> 00:17:10.260 for a Hayward Fault – the same rupture model, 00:17:10.260 --> 00:17:13.940 same Earth model – different resolutions. 00:17:13.940 --> 00:17:18.860 And previous calculations have really gone up to about 1 hertz. 00:17:18.870 --> 00:17:25.910 And a lot of hybrid simulations that use 3D methods will – might only run those 00:17:25.910 --> 00:17:30.820 simulations up to 1 hertz and then use either 1D Green’s functions, you know, 00:17:30.820 --> 00:17:35.820 from a 1D model, that don’t include basin and 3D wave propagation effects, 00:17:35.820 --> 00:17:40.620 or if they do, they only include them empirically through some scaling. 00:17:40.620 --> 00:17:43.960 Or they use the stochastic method to extend – you know, 00:17:43.960 --> 00:17:49.900 basically random numbers to extend the frequency content. 00:17:49.900 --> 00:17:53.130 What we’re able to do – what we’re able to do now is 00:17:53.130 --> 00:17:56.289 to fully calculate these in a 3D model. 00:17:56.289 --> 00:17:59.660 So, you know, with the wave propagation through the basins 00:17:59.660 --> 00:18:03.330 and the basin edges and all the mode conversions and scattering. 00:18:03.330 --> 00:18:07.289 So then each of these records, for a site in Oakland and 00:18:07.289 --> 00:18:12.309 a site in Livermore, represent the doubling of the frequency. 00:18:12.309 --> 00:18:15.510 And to double the frequency, we need to make the grid spacing 00:18:15.510 --> 00:18:18.740 in half in three dimensions and then double the number of time steps. 00:18:18.740 --> 00:18:22.600 So it’s a factor 2 to the 4th, or 16, more effort. 00:18:22.600 --> 00:18:27.490 So to go for theses four doublings, you know, it’s 65,000 times more 00:18:27.490 --> 00:18:32.929 complicated, or more effort, to calculate the 5-hertz results shown on the top than 00:18:32.929 --> 00:18:37.580 the low-frequency results on the bottom. But you can see we can get much more 00:18:37.580 --> 00:18:40.440 high-frequency energy, and the peak acceleration actually 00:18:40.440 --> 00:18:44.280 more than doubles as we go from 2-1/2 to 5 hertz. 00:18:44.280 --> 00:18:47.580 And we get a lot of high-frequency response out here 00:18:47.580 --> 00:18:51.370 in Livermore due to basin effects and such. 00:18:51.370 --> 00:18:54.600 So this is the domain that I’ll show, that we’re calculating on. 00:18:54.600 --> 00:18:59.800 It’s 120 by 80 by 30 kilometers in depth. 00:18:59.800 --> 00:19:06.430 We’ve got this rupture – this SRF. That shows the slip, rise time, and rake. 00:19:06.430 --> 00:19:10.309 It’s projected – draped onto the three-dimensional geometry 00:19:10.309 --> 00:19:14.680 of the Hayward Fault. The hypocenter, which is the green star 00:19:14.680 --> 00:19:19.920 here, is set to San Leandro, which is, you know, a bend in the Hayward Fault. 00:19:19.930 --> 00:19:25.330 And it’s important to note, too, that the Hayward Fault, on average, you know, 00:19:25.330 --> 00:19:31.080 dips to the – to the northeast. And it dips more – the dip is – well, 00:19:31.080 --> 00:19:33.520 there’s a shallower dip in the southern part of the fault 00:19:33.531 --> 00:19:35.260 compared to the northern part of the fault. 00:19:35.260 --> 00:19:38.760 And that, I think, impacts the ground motions, as we’ll see. 00:19:39.620 --> 00:19:42.980 All right. So we use the USGS 3D model that represents the, 00:19:42.990 --> 00:19:46.720 you know, Vp, Vs, rho, Qp, Qs, and the topography. 00:19:46.720 --> 00:19:50.320 It’s for a large domain – 140 by 290 kilometers. 00:19:50.320 --> 00:19:54.710 We’re only using the smaller subsection here – 120 by 80. 00:19:54.710 --> 00:19:57.610 And we’ve chose – and what I’ll show today, we’ve chose to limit the shear 00:19:57.610 --> 00:20:03.380 wave speed to 500 meters per second. And that’s a compromise. 00:20:03.380 --> 00:20:07.900 It’s a practical compromise to be able to fit a calculation 00:20:07.900 --> 00:20:12.250 and obtain a certain frequency of resolution. 00:20:12.250 --> 00:20:15.020 If we lower that frequency, then we can’t go to as 00:20:15.020 --> 00:20:19.000 high a frequency with a given resource. 00:20:19.000 --> 00:20:26.140 So we made that choice, and I’ll show, you know, where we’re – 00:20:26.140 --> 00:20:29.630 where we’re not honoring the model in the next few slides here. 00:20:29.630 --> 00:20:36.419 But it’s important to note that the weak low-velocity, low-wave speed soils 00:20:36.419 --> 00:20:42.010 that exist in a lot of the Bay Area and where people live would respond 00:20:42.010 --> 00:20:46.010 probably non-linearly to strong ground motion if they’re near the fault. 00:20:46.010 --> 00:20:49.900 And that’s something that, you know, we’d like to – so it’s really our goal to, 00:20:49.900 --> 00:20:53.620 you know, again, try to compute high-frequency simulations with a 00:20:53.620 --> 00:20:58.580 realistic minimum shear wave speed. And it’s possible that nonlinear 00:20:58.580 --> 00:21:00.780 geomechanics would need to be included in the future. 00:21:00.780 --> 00:21:06.350 You know, and Daniel Roten and others at SCEC had included plasticity in there. 00:21:06.350 --> 00:21:09.419 I won’t talk about it today, but in a separate study, 00:21:09.419 --> 00:21:11.510 we’re starting to look at this by looking at the differences 00:21:11.510 --> 00:21:16.039 between a Vs-min of 500 and 250 meters per second. 00:21:16.040 --> 00:21:18.860 So this is the shear wave speed at the surface. 00:21:18.860 --> 00:21:23.240 You know, from zero to 3.5 kilometers per second. 00:21:23.240 --> 00:21:27.190 The ophiolite body – the San Leandro gabbro or Coast Range ophiolite 00:21:27.190 --> 00:21:32.309 shows up right along the Hayward Fault. But there’s a lot of areas here in 00:21:32.309 --> 00:21:38.330 the central Bay Area along the East Bay Flats and Santa Clara Valley and parts of 00:21:38.330 --> 00:21:41.450 the peninsula here where the wave speeds are actually quite low 00:21:41.450 --> 00:21:44.480 at the surface. They’re below 500 meters per second. 00:21:44.480 --> 00:21:49.179 The minimum shear wave speed in this model is 80 meters per second. 00:21:49.179 --> 00:21:54.950 And here, this contour, then, shows the 500 meter-per-second contour is white, 00:21:54.950 --> 00:21:58.890 and it really includes some really important places where people live. 00:21:58.890 --> 00:22:02.140 So that shows where we’re not honoring the model. 00:22:02.140 --> 00:22:08.740 And this is now – it’s the depth to the 500 meter-per-second shear wave value. 00:22:08.740 --> 00:22:12.720 And it’s as deep as 75 meters in many places. 00:22:12.720 --> 00:22:15.919 But, again, along the East Bay and Oakland and Berkeley and the, 00:22:15.919 --> 00:22:19.020 you know, Hayward, Fremont, the Santa Clara Valley, and then parts of 00:22:19.020 --> 00:22:27.259 the peninsula, there are wave speeds lower than 500 meter-per-second. 00:22:27.259 --> 00:22:30.450 And then this is the depth to the 1-kilometer-per-second interval, 00:22:30.450 --> 00:22:33.740 and it just – you start to see very big differences along the Hayward Fault 00:22:33.740 --> 00:22:38.230 with the green – or the Great Valley Sequence sedimentary rocks have lower 00:22:38.230 --> 00:22:42.980 wave speed than the Franciscan rocks. So you have a thin veneer of low wave 00:22:42.980 --> 00:22:48.580 speeds west of the Hayward Fault and then a rapid increase in the Franciscan 00:22:48.580 --> 00:22:51.299 rocks here west of the Hayward Fault. 00:22:51.299 --> 00:22:53.660 But east of the Hayward Fault, you know, the wave speeds start at 00:22:53.660 --> 00:22:59.060 about 500 meters per second, but they stay lower at a much deeper depth. 00:22:59.060 --> 00:23:05.480 And then here’s the depth to the 2.5-kilometer-per-second interval. 00:23:05.490 --> 00:23:10.059 And you can see those – you know, these 2.5-meter-per-second – you know, 00:23:10.060 --> 00:23:14.240 you don’t cross that until you get to almost 4-kilometer depth, 00:23:14.240 --> 00:23:17.370 and then you’re in the East Bay Hills and Orinda, Lafayette, Moraga, 00:23:17.370 --> 00:23:22.570 et cetera. Whereas, it’s much shallower west of the Hayward Fault. 00:23:22.570 --> 00:23:26.580 And in terms of profiles, here are three profiles on different scales. 00:23:26.580 --> 00:23:31.240 The geotechnical – and this is for a site in Oakland – is in blue. 00:23:31.240 --> 00:23:35.470 And in Orinda, which is east of the Hayward Fault, in cyan. 00:23:35.470 --> 00:23:40.320 And you can see that – you know, and here’s the 500-meter-per-second level. 00:23:40.320 --> 00:23:42.550 We’re not honoring the model here when we use 00:23:42.550 --> 00:23:45.110 this 500-meter-per-second – you know, in places in Oakland. 00:23:45.110 --> 00:23:48.900 And we’re investigating what the consequences are there. 00:23:48.900 --> 00:23:53.600 But, as – but east of the Hayward Fault, in Orinda, 00:23:53.600 --> 00:23:56.500 you know, we’re able to honor the model. 00:23:56.500 --> 00:23:59.549 If we look at the upper crustal scale down to 12 kilometers, it’s not until – 00:23:59.549 --> 00:24:03.000 you know, you can see there are very large differences between the two 00:24:03.000 --> 00:24:07.250 profiles that, not surprisingly, result in different ground motions. 00:24:07.250 --> 00:24:09.549 And you have to get to depths on the crustal scale – you have to 00:24:09.549 --> 00:24:14.140 get to depths of almost 10 kilometers before those two profiles merge. 00:24:15.710 --> 00:24:19.280 So with the Franciscan wave speeds at depth being 00:24:19.299 --> 00:24:24.190 larger than the Great Valley – so west being faster than the east. 00:24:24.190 --> 00:24:29.080 And then just showing this again here for the surface values. 00:24:29.080 --> 00:24:33.440 All right, so last summer, we ran on the Cori machine 00:24:33.440 --> 00:24:38.750 this 120-by-80-kilometer domain. We ran without mesh refinement 00:24:38.750 --> 00:24:42.420 in an earlier version of the code, and we were able to get to 4.2 hertz. 00:24:42.420 --> 00:24:44.700 And that was using about two-thirds of that machine. 00:24:44.700 --> 00:24:47.690 And that was quite a herculean effort at the time. 00:24:47.690 --> 00:24:51.230 That was the highest resolution calculation that had been done. 00:24:51.230 --> 00:24:55.159 But then we did it a few months later on the Quartz machine with a version 00:24:55.159 --> 00:24:58.039 of SW4 that had mesh refinement. 00:24:58.040 --> 00:25:01.480 And mesh refinement – just to illustrate how important it is, 00:25:01.480 --> 00:25:06.540 we went from 87 billion grid points for a fixed grid to 14 – 00:25:06.540 --> 00:25:09.971 almost 15 billion with mesh refinement. 00:25:09.971 --> 00:25:14.309 So we have six times fewer points. And so mesh refinement is 00:25:14.309 --> 00:25:18.210 really important for us to be able to, you know, make the calculations 00:25:18.210 --> 00:25:21.040 smaller so they’ll fit on a given resource. 00:25:21.040 --> 00:25:25.240 And we published this in a paper in GRL that came out in January. 00:25:25.240 --> 00:25:29.260 Then, in June, instead of here, you know, we were working on 00:25:29.260 --> 00:25:36.990 getting ready for this 5-hertz calculation with 26 billion grid points – 26 billion 00:25:36.990 --> 00:25:41.230 grid points using mesh refinement on almost all of the Cori machine. 00:25:41.230 --> 00:25:45.040 And that – at the time, that was the highest resolution. 00:25:45.040 --> 00:25:49.640 And we started writing a paper on that, and we submitted that to SRL. 00:25:49.640 --> 00:25:54.060 And I will show – let’s see here. 00:25:54.920 --> 00:25:58.900 Yeah. There’s an animation of that here. 00:26:00.760 --> 00:26:04.080 So this is going to show – there’s the PGV map that results from this 00:26:04.080 --> 00:26:08.840 5-hertz calculation. And you can see the asymmetries across the Hayward Fault. 00:26:08.840 --> 00:26:14.260 So this is now – it’s a 77-kilometer-long bilateral rupture. 00:26:14.260 --> 00:26:17.100 It starts here in San Leandro. 00:26:17.100 --> 00:26:20.640 And this is for the 3D calculation with topography. 00:26:20.650 --> 00:26:22.840 So this is showing the magnitude of ground velocity. 00:26:22.840 --> 00:26:25.970 So we take just the vectors – you know, the sum of the squares – 00:26:25.970 --> 00:26:29.760 square to the sum of the squares of the three components of ground motion, 00:26:29.760 --> 00:26:31.710 and that’s the magnitude of the velocity shown in 00:26:31.710 --> 00:26:35.020 something like a color – a ShakeMap color palette. 00:26:35.020 --> 00:26:38.350 And you can see there are some large amplitudes here 00:26:38.350 --> 00:26:40.490 east of the Hayward Fault. 00:26:40.490 --> 00:26:44.840 And you can start to see two that the waves are maybe moving out 00:26:44.840 --> 00:26:48.610 faster on the western side of the Hayward Fault 00:26:48.610 --> 00:26:52.030 and delayed on the eastern side of the Hayward Fault. 00:26:52.030 --> 00:26:53.610 And you can see there’s variation. 00:26:53.610 --> 00:26:58.210 Now that Coast Range ophiolite – that high-velocity body right along the 00:26:58.210 --> 00:27:03.160 Hayward Fault acts as a shield, and it has low motions because it’s so stiff. 00:27:04.260 --> 00:27:08.680 And then there’s, of course, variations in the slip across – along the fault. 00:27:08.690 --> 00:27:11.120 So that gives rise to variations in the ground motion. 00:27:11.120 --> 00:27:13.809 But there are asymmetries. We see larger ground motions here 00:27:13.809 --> 00:27:17.050 on the eastern side of the Hayward Fault compared to the west. 00:27:17.050 --> 00:27:19.080 And now we really start to see some 3D effects. 00:27:19.080 --> 00:27:21.600 You can see waves entering into the Evergreen Basin and 00:27:21.600 --> 00:27:27.140 are delayed and amplified here along the southern end of the fault. 00:27:27.140 --> 00:27:30.090 And then the – you know, the northern extension of the 00:27:30.090 --> 00:27:33.140 Hayward Fault into the Rodgers Creek Fault across San Pablo Bay. 00:27:33.140 --> 00:27:37.120 Well, there’s debate about the exact geometry of that, I’m sure. 00:27:37.120 --> 00:27:41.669 But we can certainly see that the waves are moving out faster on the west side, 00:27:41.669 --> 00:27:46.250 and they’re slower on the east side and amplified. 00:27:46.250 --> 00:27:49.179 And that really shows up as we progress. 00:27:49.179 --> 00:27:52.940 Then we see these things like the San Leandro Basin offshore here 00:27:52.940 --> 00:27:57.730 in the bay show up in the Cupertino and Evergreen Basins. 00:27:57.730 --> 00:28:01.360 And the San Pablo Bay. The Napa Valley show up. 00:28:01.360 --> 00:28:04.500 And then the Dublin/Pleasanton Valley actually shows up a lot. 00:28:04.500 --> 00:28:08.480 And the hard rock here in the Diablo Range kind of reflects 00:28:08.480 --> 00:28:12.490 energy back into the Livermore Valley. 00:28:13.460 --> 00:28:17.880 And this calculation was run for 90 seconds of seismogram time. 00:28:17.899 --> 00:28:23.470 So – and then we can actually see, you know, some very late-arriving 00:28:23.470 --> 00:28:28.060 pulses out there in some of the basins – very slow waves. 00:28:31.220 --> 00:28:32.700 And on and on. 00:28:32.820 --> 00:28:37.200 And then, by the end, it’s just a lot of coda bouncing around. 00:28:37.780 --> 00:28:39.600 All right. 00:28:41.560 --> 00:28:49.000 So that’s the 5-hertz run. This is just a snapshot from that. 00:28:49.010 --> 00:28:55.240 So we like to run, for a given rupture, both a 1D flat model without surface 00:28:55.240 --> 00:28:59.310 topography along with the 3D model. And the PGV maps from those 00:28:59.310 --> 00:29:03.630 two calculations are shown here – 1D flat and 3D topography. 00:29:03.630 --> 00:29:09.580 Now, because the – we’re honoring the fault geometry, and also because 00:29:09.580 --> 00:29:14.820 of rake variations that are part of the Graves and Pitarka method, there are – 00:29:14.820 --> 00:29:17.080 we don’t get perfect symmetry across the Hayward Fault. 00:29:17.080 --> 00:29:20.750 And you can see there’s a lot of asymmetry with the eastern side 00:29:20.750 --> 00:29:23.730 of the Hayward Fault toward – the direction toward which the fault 00:29:23.730 --> 00:29:27.760 is dipping have larger motions than the western side. 00:29:27.760 --> 00:29:34.940 And that’s mostly in the south where the dip is strongest – or, the shallowest. 00:29:34.940 --> 00:29:39.360 But towards the north, there’s more symmetry across the fault. 00:29:39.360 --> 00:29:43.020 And so we’re trying to look into that, you know, and really quantify 00:29:43.029 --> 00:29:47.620 how much of the variability that we see across the fault 00:29:47.620 --> 00:29:50.880 is due to fault geometry and how much is due to 3D structure. 00:29:50.880 --> 00:29:55.260 Here, we see much more asymmetry with this northern part of the 00:29:55.260 --> 00:29:58.380 Hayward Fault here – areas in El Sobrante and Orinda 00:29:58.380 --> 00:30:02.150 having very high motions compared to, say, Berkeley and Richmond. 00:30:02.150 --> 00:30:05.700 And the Coast Range ophiolite shows up, again, with low motions 00:30:05.700 --> 00:30:07.700 because of the hard rock there. 00:30:07.700 --> 00:30:13.060 And you can see, you know, some of these hot spots are due to slip, 00:30:13.060 --> 00:30:16.620 you know, on the fault, but also the lower wave speeds that exist in 00:30:16.620 --> 00:30:21.020 those sedimentary deposits in the very top of the Franciscan. 00:30:21.020 --> 00:30:25.490 But it’s nice to do this 1D and 3D because we can take the 3D solution 00:30:25.490 --> 00:30:30.600 and essentially divide out the 1D solution and eliminate the source effects. 00:30:30.600 --> 00:30:35.490 So we’re then getting the motions at a given point from the 3D model 00:30:35.490 --> 00:30:39.780 can be normalized by what would be experienced with a 1D model. 00:30:39.780 --> 00:30:44.260 And we’ve done that. I’ll talk about that a little bit later. 00:30:44.260 --> 00:30:47.231 But let’s just look here now at – these are ground motion intensity 00:30:47.231 --> 00:30:52.990 measures, as a function of distance, with the Abrahamson, Silva, and Kamai 00:30:52.990 --> 00:30:59.820 2014 ground motion model GMM. So GMIMs, you know, with GMMs. 00:30:59.820 --> 00:31:05.430 For a 1D flat on the left versus the 3D topographic case on the right. 00:31:05.430 --> 00:31:08.850 And this is PGA – peak ground acceleration 00:31:08.850 --> 00:31:14.549 from 100 meters, Joyner distanced out to 60 kilometers. 00:31:14.549 --> 00:31:22.180 And they’re color-coded by the – essentially the fault-normal distance – 00:31:22.180 --> 00:31:23.800 distance from the fault. 00:31:23.809 --> 00:31:30.490 Now, what we see is good agreement with the – with the model. 00:31:30.490 --> 00:31:34.600 So the median prediction is shown as the solid line, and the 00:31:34.600 --> 00:31:41.360 one-sigma multiply/divide uncertainties are given as the dashed lines. 00:31:41.360 --> 00:31:44.550 And in each of these plots, we have the mean value of the residuals 00:31:44.550 --> 00:31:49.649 relative to that using the site-specific parameters for each – 00:31:49.649 --> 00:31:52.179 for each site and the standard deviation. 00:31:52.179 --> 00:31:56.010 So we’re getting, you know, 0.4 log units of standard deviation. 00:31:56.010 --> 00:32:03.070 And that’s less than the, say, 0.7 or so of the error in the model. 00:32:03.070 --> 00:32:06.510 And the median in this case, for the 1D model, it’s 0.5. 00:32:06.510 --> 00:32:12.130 You know, we’re low by 0.5 log units for the 1D model. 00:32:12.130 --> 00:32:15.950 For the 3D model, we’re about 0.3 log units low. 00:32:15.950 --> 00:32:19.740 And what we consistently see is that the 3D model produces 00:32:19.740 --> 00:32:23.529 larger variations – larger uncertainties than what – 00:32:23.529 --> 00:32:27.010 larger scatter than what’s seen in the 1D model. 00:32:27.010 --> 00:32:31.350 So that’s PGA. And, you know, this is exciting to me that we’re able 00:32:31.350 --> 00:32:36.480 to match PGA, you know, with a 5-hertz calculation, fairly well. 00:32:36.480 --> 00:32:40.940 And then, if we go to PGV, the median prediction for the 1D model is 00:32:40.950 --> 00:32:46.730 pretty much zero with a variance of – or, a standard deviation of 0.3 and 00:32:46.730 --> 00:32:50.400 0.1 for the 3D model, but about half as much, 00:32:50.400 --> 00:32:57.680 you know, variation in terms of more scatter in the 3D model. 00:32:57.690 --> 00:33:02.970 All right, so that was all going along a couple – earlier in the summer, though, 00:33:02.970 --> 00:33:09.740 we were asked if we could run – would be able to run on the Sierra machine. 00:33:09.740 --> 00:33:13.809 So we’d been running on Cori. A few years ago, this was higher up 00:33:13.809 --> 00:33:18.309 on the top 100 list of the – you know, IEEE’s fastest computers. 00:33:18.309 --> 00:33:25.200 It’s a 27-petaflop system. It uses 68 CPU cores per node. 00:33:25.200 --> 00:33:28.100 And that’s at Lawrence Berkeley lab. 00:33:28.110 --> 00:33:30.940 And that’s a – you know, the ECP, a big of that was to be able to 00:33:30.940 --> 00:33:38.220 run efficiently on that architecture. And we showed that in June. 00:33:38.220 --> 00:33:42.240 But meanwhile, Livermore had been – had a small version of this 00:33:42.240 --> 00:33:45.179 new computer, Sierra, which is a different architecture. 00:33:45.179 --> 00:33:50.380 It has four GPU cards per node. 00:33:50.380 --> 00:33:54.080 And it’s, you know, planned to be a – over 100-petaflop machine. 00:33:54.080 --> 00:33:56.760 It’s currently number 3 on the top 100. 00:33:56.760 --> 00:34:01.299 So Livermore interval investment was worked to port SW4 to the 00:34:01.299 --> 00:34:06.970 GPU platform using some software that means you don’t have to go to the very 00:34:06.970 --> 00:34:13.609 low-level GPU programming that used to be required in the early days of GPUs. 00:34:13.609 --> 00:34:20.450 So we ended up running the 5-hertz case that we did on Cori 00:34:20.450 --> 00:34:25.550 on Sierra a couple of weeks ago, and that went really well. 00:34:25.550 --> 00:34:28.869 The important thing to note is that this is where, you know, next-generation 00:34:28.869 --> 00:34:33.859 computing is going. It’s going to be heavily dependent on GPUs. 00:34:33.860 --> 00:34:39.360 So we ran, again, the 5-hertz – we did a 5-hertz run on Cori. 00:34:39.360 --> 00:34:41.760 And that was the – you know, 26 billion grid points, 00:34:41.760 --> 00:34:45.129 you know, 63,000 time steps. We used checkpointing, so, you know, 00:34:45.129 --> 00:34:48.530 we wrote out the solution at two adjacent time steps 00:34:48.530 --> 00:34:52.659 every 4,000 time steps, so if it crashed, we could restart. 00:34:52.660 --> 00:34:56.380 And that capability has now been mapped over to the GPU version. 00:34:56.380 --> 00:34:59.720 And that took 10 hours on 85% of Cori. 00:34:59.730 --> 00:35:04.681 On the Sierra system, we had the same mesh, the same – a lot of the 00:35:04.681 --> 00:35:09.560 same details, but a finer grid spacing that allowed us to get to 6.9 hertz. 00:35:09.560 --> 00:35:14.110 And that was then 68 billion grid points and 87,000 time steps. 00:35:14.110 --> 00:35:17.730 No checkpointing in the version we ran about two weeks ago. 00:35:17.730 --> 00:35:21.740 But that was only using 25% of the Sierra machine and a little bit longer – 00:35:21.740 --> 00:35:24.290 in 12 hours. So we now have results for 00:35:24.290 --> 00:35:29.671 a fully deterministic 3D simulation of a magnitude 7 earthquake 00:35:29.671 --> 00:35:33.190 on the Hayward Fault with 3D, you know, geology from the USGS 00:35:33.190 --> 00:35:36.839 and the surface topography and attenuation. 00:35:36.839 --> 00:35:39.960 And this is a snapshot from that. I don’t have an animation of this. 00:35:39.960 --> 00:35:45.720 But we’re resolving, in this case, waves that are on – as short as 100 meters. 00:35:45.720 --> 00:35:49.260 And so there’s some finer detail that’s in there. 00:35:49.270 --> 00:35:52.270 And here’s the ShakeMap that results from that, and we do get some larger 00:35:52.270 --> 00:35:57.630 motions in the near-fault area. But a lot of the features that we saw are the same. 00:35:57.630 --> 00:36:01.120 You know, the Coast Range ophiolite is a seismic shield. 00:36:01.120 --> 00:36:04.400 And the, you know, eastern – the northern part of the Hayward Fault, 00:36:04.410 --> 00:36:10.950 we see large motions in El Sobrante and Orinda and Moraga compared to over in 00:36:10.950 --> 00:36:16.340 the East Bay in Richmond and El Cerrito and Berkeley and Oakland, et cetera. 00:36:17.030 --> 00:36:18.760 So this is what we get in terms of the wave fields. 00:36:18.760 --> 00:36:23.960 I showed this before, but I didn’t – now we can include the 6.9 hertz. 00:36:23.960 --> 00:36:26.960 And we see almost a – you know, it’s not quite doubling the frequency, 00:36:26.960 --> 00:36:31.550 but we’re certainly doubling the peak acceleration that, you know, 00:36:31.550 --> 00:36:34.810 is associated with this peak here at the site in Oakland. 00:36:34.810 --> 00:36:40.220 And, in Livermore, you know, we are seeing, you know, larger motions. 00:36:40.220 --> 00:36:43.380 They’re almost doubling the peak acceleration. 00:36:43.380 --> 00:36:47.180 And then, to just look at these sites again – so we output – 00:36:47.180 --> 00:36:51.650 on a 2-kilometer grid, we output the time histories. 00:36:51.650 --> 00:36:57.151 And the X is essentially fault-parallel, Y is fault-normal, and Z is vertical. 00:36:57.151 --> 00:37:00.079 And here’s the accelerations. You know, we’re about half a g 00:37:00.079 --> 00:37:04.930 for a site in Oakland on top of the fault. Peak – and then velocity is here, 00:37:04.930 --> 00:37:08.609 maybe 30-centimeters-per-second peak. And there’s a strong velocity pulse 00:37:08.609 --> 00:37:12.430 in here that you can’t see when you look at the whole 90 seconds. 00:37:12.430 --> 00:37:15.260 And then, in terms of displacements, we see the displacement step 00:37:15.260 --> 00:37:16.950 in the permanent deformation. 00:37:16.950 --> 00:37:21.290 So we’re resolving the low frequencies up to the highest frequency. 00:37:21.290 --> 00:37:25.440 In terms of response spectra, now the simulation produces 00:37:25.440 --> 00:37:28.880 the RotD50 response spectra that’s shown in black. 00:37:28.880 --> 00:37:34.040 And that’s compared with four of the NGA-West2 GMPEs, shown in color, 00:37:34.050 --> 00:37:36.660 with their uncertainties as the dashed lines. 00:37:36.660 --> 00:37:43.630 This is the full range of the GMPEs from 100 hertz, or 0.01 second, all the 00:37:43.630 --> 00:37:51.090 way up to 10 seconds. So from 0.1 hertz to, you know, 1 hertz, to 10 hertz here. 00:37:51.090 --> 00:37:55.430 So the gray area is what we’re not resolving well in the simulation. 00:37:55.430 --> 00:37:58.890 And then these are the site-specific, you know, GMPE parameters 00:37:58.890 --> 00:38:03.300 for this particular site. And we get good agreement with – 00:38:03.300 --> 00:38:07.140 you know, and all the way out with the – with the 00:38:07.150 --> 00:38:10.230 ground motion prediction equation, or ground motion models. 00:38:10.230 --> 00:38:14.880 And even the PGA – the spectral ordinate looks pretty good as well. 00:38:14.880 --> 00:38:17.440 For a site in Livermore, a much more complex response – 00:38:17.440 --> 00:38:22.230 very long duration scattered wave field that you saw in the animation – 00:38:22.230 --> 00:38:24.980 smaller permanent displacements, big, you know, velocity pulse 00:38:24.980 --> 00:38:27.240 from the – from the direct waves. 00:38:27.240 --> 00:38:30.099 Some directivity may enter into this. 00:38:30.100 --> 00:38:35.080 In terms of response spectra, again, this is from 6.9 hertz. 00:38:35.080 --> 00:38:39.660 Shortest period is about – what is that – 0.15 hertz 00:38:39.661 --> 00:38:41.810 or seconds or something like that. 00:38:41.810 --> 00:38:48.010 And we have somewhat larger response at this site, which could be 00:38:48.010 --> 00:38:52.400 attributed to directivity and unmodeled basin effects 00:38:52.400 --> 00:38:57.600 that aren’t encoded in the GMPE – so the ground motion models. 00:38:57.600 --> 00:38:59.600 Then – that’s for a couple sites. 00:38:59.600 --> 00:39:04.839 Now, on aggregate, here this is all of the sites from 100 meters 00:39:04.839 --> 00:39:11.060 out to 60 kilometers with the – this is PGA – peak ground acceleration. 00:39:11.060 --> 00:39:14.890 And the mean is essentially zero. So this is really exciting to me that 00:39:14.890 --> 00:39:18.740 we’re able to simulate PGA with a fully deterministic 00:39:18.740 --> 00:39:22.579 simulation all the way up to 6.9 hertz. 00:39:22.579 --> 00:39:27.579 And, you know, the standard deviation is 0.5 or, you know, on the order, or less, 00:39:27.579 --> 00:39:31.440 than the – than the reported value in the ground motion models. 00:39:31.440 --> 00:39:33.340 Here’s PGV. 00:39:33.340 --> 00:39:36.150 And, again, very small bias. 00:39:36.150 --> 00:39:40.460 And then here’s, you know, 10 hertz or 1.1 seconds. 00:39:40.460 --> 00:39:46.290 So even, you know, beyond where we’re resolving, beyond 6.9 hertz, we’re 00:39:46.290 --> 00:39:50.070 able to match the spectral accelerations and the ground motion models. 00:39:50.070 --> 00:39:52.600 But there’s a lot – a lot of scatter. 00:39:52.600 --> 00:39:58.400 And then here is 5 hertz, or 0.2 seconds, 0.3 seconds, 3.3 hertz. 00:39:58.400 --> 00:40:04.480 0.5 seconds – 2-hertz waves here. And then 1 hertz. 00:40:04.480 --> 00:40:07.340 And now going to 3-second periods – the longer periods. 00:40:07.349 --> 00:40:13.980 Somewhat – a little bit – this was biased high by, you know, 0.5 or so log units. 00:40:13.980 --> 00:40:17.020 And then now let’s look at bias across the band. 00:40:17.020 --> 00:40:22.270 So this is the 1D case resolved to 5 hertz. And then this is the bias. 00:40:22.270 --> 00:40:26.230 So we take – for all of the 2,300 sites, 00:40:26.230 --> 00:40:28.650 we take the measured ground motion intensity. 00:40:28.650 --> 00:40:34.500 We divide by the site-specific ground motion prediction equation estimate. 00:40:34.500 --> 00:40:37.080 We form the ratio, and we take the logarithm of that. 00:40:37.080 --> 00:40:39.450 And if we had a perfect simulation that agreed 00:40:39.450 --> 00:40:42.730 with the GMPEs, we’d be at zero. 00:40:42.730 --> 00:40:46.950 And then, we – so each of these bars represents the – 00:40:46.950 --> 00:40:51.440 all the sites that were recorded across the domain with a median value 00:40:51.440 --> 00:40:56.319 as the thick line, and then the box-and-whisker plots here. 00:40:56.319 --> 00:41:02.250 So this is the – 50% of the data are within the cyan boxes, 00:41:02.250 --> 00:41:07.760 and then the 1.5 times that inter-quartile range is indicated by the bars, 00:41:07.760 --> 00:41:10.580 and then outliers are shown as the circles. 00:41:10.580 --> 00:41:15.040 Those are individual measurements that exceed 1.5 times the inter-quartile range. 00:41:15.040 --> 00:41:17.900 So, over here, we get good agreement. 00:41:17.900 --> 00:41:22.310 And for reference, I’ve drawn in 0.7 log units with the red dashed lines as a 00:41:22.310 --> 00:41:26.609 nominal one-sigma uncertainty of the ground motion models. 00:41:26.609 --> 00:41:35.020 So, in this case, for 5 hertz, 1D model, PGA is on the edge of being within 00:41:35.020 --> 00:41:38.220 the uncertainties of the ground motion models, being near zero. 00:41:38.220 --> 00:41:41.630 But all of the spectral – RotD50 spectral accelerations 00:41:41.630 --> 00:41:50.800 are within the errors to about 4 seconds, or 0.4 – 4 hertz, or 0.25 seconds. 00:41:50.800 --> 00:41:54.840 These over here in the gray area are poorly resolved. 00:41:55.400 --> 00:41:58.540 And that’s for the 1D model. For the 3D model, we get this. 00:41:58.540 --> 00:42:01.450 So you can see a lot more scatter, so it’s very clear that the 00:42:01.450 --> 00:42:04.160 3D model produces more scatter than the 1D model. 00:42:04.160 --> 00:42:08.890 Of course, the ground motion – the rupture generators are often calibrated 00:42:08.890 --> 00:42:14.970 with 1D models, so we better get good agreement across the band here. 00:42:14.970 --> 00:42:19.180 But we do, in the 3D model, see more variation. 00:42:19.180 --> 00:42:25.720 And, you know, more variation in terms of the spread of the data. 00:42:25.730 --> 00:42:28.420 And then, for the 6.9-hertz case, this is what we get. 00:42:28.420 --> 00:42:32.970 So now we’re almost resolving within the GMPE errors 00:42:32.970 --> 00:42:36.710 all the way out to 10 hertz. So this is now for two whole decades 00:42:36.710 --> 00:42:41.280 of frequency, or period, which is really exciting to me. 00:42:42.120 --> 00:42:45.300 All right. So, again, I talked about the ratios, 00:42:45.300 --> 00:42:49.740 where there’s the 1DFLAT to the 3DTOPO back to the 5-hertz run. 00:42:49.740 --> 00:42:55.310 If we then take the ratio of those PGV maps, we get this. 00:42:55.310 --> 00:43:00.550 So this shows PGV at a given location, and it shows the ratio. 00:43:00.550 --> 00:43:05.710 And, again, it’s the log of the ratio of the 3D divided by the 1D. 00:43:05.710 --> 00:43:08.920 So it shows us how much of the amplification in red, 00:43:08.920 --> 00:43:12.760 or de-amplification in blue, can be attributed to 3D propagation 00:43:12.760 --> 00:43:15.730 that arises as path and site effects. 00:43:15.730 --> 00:43:19.960 So what we see is, along this Coast Range ophiolite, you know, the blue, 00:43:19.960 --> 00:43:25.400 it’s actually faster than the 1D model, so we get de-amplification in the 3D model. 00:43:25.400 --> 00:43:28.920 And we see that also in the Diablo Range, and in Mount Diablo, and the 00:43:28.920 --> 00:43:34.250 hills between the Great Valley and Napa Valley, and, to some extent, Sonoma. 00:43:34.250 --> 00:43:38.369 The Marin Headlands and some of the hard rock 00:43:38.369 --> 00:43:42.040 associated with the hills here in the South Bay. 00:43:42.040 --> 00:43:46.040 In the basins, we see the red associated with amplification. 00:43:46.040 --> 00:43:51.400 And there’s the San Leandro Basin and areas near the Hayward Fault, 00:43:51.400 --> 00:43:55.490 where we have low wave speeds. And then very much so where we 00:43:55.490 --> 00:43:59.190 have the thick Great Valley Sequence. In fact, all of the East Bay – 00:43:59.190 --> 00:44:04.690 Livermore Valley, the delta, San Pablo Bay – shows elevated motions. 00:44:04.690 --> 00:44:08.480 And also in the Golden Gate and areas – 00:44:08.480 --> 00:44:13.740 the headlands west of the San Andreas Fault, it looks like. 00:44:14.460 --> 00:44:18.050 We did the same thing with a different rupture model with a slightly different 00:44:18.050 --> 00:44:24.460 domain in – with the 4-hertz calculation, and we got this map, which – it’s a 00:44:24.460 --> 00:44:27.580 slightly different domain, but we still see the Coast Range ophiolite, you know, 00:44:27.580 --> 00:44:30.000 Mount Diablo – this doesn’t have the topographic shading 00:44:30.000 --> 00:44:38.119 in there like the one – but we see similar patterns of this ratio. 00:44:38.120 --> 00:44:42.880 So that’s encouraging because it means that we can – we can hope to 00:44:42.880 --> 00:44:48.840 isolate path and site effects by forming these ratios of 3D to 1D. 00:44:48.840 --> 00:44:54.020 And then, for the – this one is always slow to – okay, this is – 00:44:54.020 --> 00:45:00.730 so, for PGV, for the 4-hertz calculations that’s described in our GRL paper, 00:45:00.730 --> 00:45:05.740 and the 5-hertz calculation on the right, which I’ve just been talking about, 00:45:05.740 --> 00:45:08.859 this is the PGV maps, or the ratio maps, 00:45:08.860 --> 00:45:14.040 that we get from individual seismograms that are output on a 2-kilometer grid. 00:45:14.040 --> 00:45:18.180 And they show a similar pattern here where the East Bay – 00:45:18.180 --> 00:45:22.369 here, the fault trace is shown here. This trace fault is 50 kilometers long. 00:45:22.369 --> 00:45:27.470 This is 75 – or, 77 kilometers long. But we see, east of the Hayward Fault, 00:45:27.470 --> 00:45:30.310 we get red values associated with amplified motion in 00:45:30.310 --> 00:45:34.020 a 3D model relative to the 1D. Along the Coast Range ophiolite, 00:45:34.020 --> 00:45:37.400 we see low wave speeds, and then, in the South Bay here – 00:45:37.400 --> 00:45:42.340 and this domain has moved further south here. 00:45:42.340 --> 00:45:49.680 And I made these maps of the depth to the shear wave speed of 1 kilometer. 00:45:49.680 --> 00:45:53.370 That shows the model, and you can see the East Bay Hills here have the 00:45:53.370 --> 00:45:55.460 low wave speed – the Great Valley Sequence, 00:45:55.460 --> 00:45:59.030 and the San Leandro Basin and the Santa Clara Valley shown here. 00:45:59.030 --> 00:46:01.460 And there’s some visual correlation here 00:46:01.460 --> 00:46:06.160 between the red values and the deeper sedimentary. 00:46:06.160 --> 00:46:11.609 And to investigate that, we took those ratios and plotted them 00:46:11.609 --> 00:46:17.240 as a function of the site-specific ground motion prediction equation values – 00:46:17.240 --> 00:46:22.540 Vs30, Z-1.0 and Z-2.5. And, you know, there’s a lot of – 00:46:22.540 --> 00:46:27.120 there’s some cases – a lot of cases that don’t show great correlation, 00:46:27.120 --> 00:46:30.780 especially at higher frequencies. But PGV and spectral acceleration 00:46:30.780 --> 00:46:35.980 at 2 hertz – or, 2 seconds, they do show correlations. 00:46:35.980 --> 00:46:40.730 And some of the longer-period ground motion intensity measures 00:46:40.730 --> 00:46:43.940 do show correlations. And it shows that, as you increase 00:46:43.940 --> 00:46:49.670 Vs30 from, say, 500 to 1,200, you decrease the motion in the 00:46:49.670 --> 00:46:55.500 3D model relative to the 1D model. And the – so the points are for, 00:46:55.500 --> 00:46:59.359 you know, individual measurements made in – 00:46:59.359 --> 00:47:03.349 from the 3D model relative to the 1D model in the simulation. 00:47:03.349 --> 00:47:10.000 And the blue shows the correlation – or, the regression analysis, and the 00:47:10.000 --> 00:47:14.170 one-sigma errors, in dashed, that are associated with the simulations. 00:47:14.170 --> 00:47:17.569 And the red show what you would get from the Abrahamson, Silva, 00:47:17.569 --> 00:47:21.320 and Kamai 2014 ground motion prediction equation. 00:47:21.320 --> 00:47:24.680 And we get some consistency between those. 00:47:24.680 --> 00:47:28.640 And then – here is for Z-1.0 and Z-2.5. 00:47:28.640 --> 00:47:33.220 And, in these cases, there are, you know, the right trends, 00:47:33.230 --> 00:47:39.010 as the basins get deeper, Z-1.0 increases for a site on top of a basin. 00:47:39.010 --> 00:47:43.089 We get increased motion in the 3D model relative to the 1D model. 00:47:43.089 --> 00:47:46.710 So this is encouraging that some of the physics – you know, the physics that’s 00:47:46.710 --> 00:47:52.680 in the simulation is able to reproduce some of the empirical behavior that’s 00:47:52.680 --> 00:47:57.680 in the ground motion equations. What we really would like to do is know 00:47:57.680 --> 00:48:02.500 why is this particular point at that value? And what is it about, you know, 00:48:02.500 --> 00:48:06.760 the – that particular site or the path that the energy has taken from the – 00:48:06.760 --> 00:48:12.839 you know, the source to the receiver that, you know, increases the motion? 00:48:12.839 --> 00:48:15.751 And that’s kind of where non-ergodic ground motion 00:48:15.751 --> 00:48:19.339 prediction equations are coming – I mean, are going to. 00:48:19.340 --> 00:48:25.190 And this is how 3D simulations can help advance hazard. 00:48:26.380 --> 00:48:29.760 All right, so to summarize, you know, the Hayward Fault, you know, provides 00:48:29.760 --> 00:48:34.080 a meaningful study area for modeling large and damaging earthquakes. 00:48:34.080 --> 00:48:37.940 There’s high hazard in an urban area in the San Francisco Bay area, 00:48:37.940 --> 00:48:44.829 especially the East Bay. With 150th anniversary of the 1868 event, it’s a 00:48:44.829 --> 00:48:49.630 great opportunity for public outreach and for people to know about this. 00:48:49.630 --> 00:48:54.859 Our recent efforts have enabled increased resolution and shorter runtimes 00:48:54.859 --> 00:48:59.660 for regional-scale motions and simulations, enabling us to do – 00:48:59.660 --> 00:49:04.440 to do more on the available resources. And this has really been made possible 00:49:04.440 --> 00:49:08.970 with advances in SW4 numerical and computational, as well as – 00:49:08.970 --> 00:49:13.650 both for the ECP, but also to be able to run on GPU systems and 00:49:13.650 --> 00:49:18.320 access to world-class computing at the DOE labs that we’ve been working at. 00:49:18.320 --> 00:49:21.160 And our computed motions are – they show good agreement with ground 00:49:21.160 --> 00:49:27.140 motion models across a broad frequency range from, you know, 0.1 to 10 hertz. 00:49:27.140 --> 00:49:30.880 Our single-event median and variance are reasonable. 00:49:30.880 --> 00:49:36.180 And the 3D Earth model results in large scatter in the ground motion intensities 00:49:36.180 --> 00:49:41.130 that, you know, I think is going to be interesting to try to work further on that. 00:49:41.130 --> 00:49:44.570 And our simulations are useful for engineering analysis, and that’s 00:49:44.570 --> 00:49:49.180 a whole nother topic that Dave or Mamun could talk about. 00:49:49.180 --> 00:49:52.450 So some caveats and needs for further investigations – next steps. 00:49:52.450 --> 00:49:55.880 There’s the USGS model. So we’d really like to thoroughly 00:49:55.880 --> 00:50:00.569 evaluate the model with real data from moderate earthquakes, try to understand, 00:50:00.569 --> 00:50:04.819 how does it perform as a function of frequency path and geologic unit. 00:50:04.819 --> 00:50:09.060 We talked a lot about this in March when Brad hosted a workshop here. 00:50:09.060 --> 00:50:13.579 You know, I like to think of the model – you know, the USGS model, 00:50:13.579 --> 00:50:18.579 the San Francisco Bay model, very – as, you know, not the model, but a model. 00:50:18.579 --> 00:50:23.559 I mean, it has defensible, you know, geometry of the surface context 00:50:23.559 --> 00:50:27.730 of geology and the, you know, overall large-scale geometry of 00:50:27.730 --> 00:50:32.800 basins is probably well-represented in the model. 00:50:32.800 --> 00:50:35.869 But it probably could be – I mean, I’m sure we’d all agree 00:50:35.869 --> 00:50:38.020 that it can be improved. 00:50:38.020 --> 00:50:41.570 And one way to do that is through a data-driven full-waveform inversion. 00:50:41.570 --> 00:50:46.300 And that’s something we hope to do as part of our ECP. 00:50:46.300 --> 00:50:52.160 And then, another topic on this is, can we really develop the model-based 00:50:52.160 --> 00:50:56.980 non-ergodic, you know, path effects GMPE to reduce scatter? 00:50:56.980 --> 00:51:01.180 Can we – can we use simulations to better understand path and site effects 00:51:01.180 --> 00:51:05.180 and use that to reduce the scatter that maps into uncertainty in the hazard, 00:51:05.180 --> 00:51:08.780 which is really, really needed? 00:51:09.480 --> 00:51:12.960 We’ve only run really two magnitude 7, you know, 00:51:12.960 --> 00:51:15.820 earthquake ruptures at this resolution. 00:51:15.820 --> 00:51:18.360 They were different segments with different slip distributions. 00:51:18.369 --> 00:51:22.530 You know, we want to look at how repeatable are our results, 00:51:22.530 --> 00:51:29.240 and what features of these simulations are repeatable and are – might be able to 00:51:29.240 --> 00:51:35.250 be determined deterministically for, say, a non-ergodic GMPE. 00:51:35.250 --> 00:51:38.170 But we want to look into how the results – you know, ground motion 00:51:38.170 --> 00:51:41.530 measures vary as a function of, you know, the hypocenter 00:51:41.530 --> 00:51:45.390 and directivity, slip distribution, rupture speed, rise time, stress drop – 00:51:45.390 --> 00:51:48.220 there’s a lot of parameters – a lot of parameter studies that can be done. 00:51:48.220 --> 00:51:51.579 We’re probably not going to do them right away at 5 hertz, but at a lower 00:51:51.579 --> 00:51:55.619 frequency using, you know, the resources we have access to. 00:51:55.619 --> 00:51:57.230 And then that would be a valuable data set 00:51:57.230 --> 00:52:01.220 for looking at path effects and site effects. 00:52:01.220 --> 00:52:03.990 And the median and overall variance that we’re getting is consistent 00:52:03.990 --> 00:52:07.329 with the ground motion models. And this is very encouraging. 00:52:07.329 --> 00:52:10.110 But we need to decompose that variance into, you know, 00:52:10.110 --> 00:52:13.280 how much is due to source, path, and site. 00:52:13.280 --> 00:52:17.060 And to do that, you know, will require us to do sort of suites 00:52:17.069 --> 00:52:21.839 of carefully designed, you know, sets of simulations to look at 00:52:21.839 --> 00:52:27.230 the within – you know, site-to-site variability, inter-event variability, 00:52:27.230 --> 00:52:31.260 single-station sigma – how much does the site response vary, 00:52:31.260 --> 00:52:35.579 you know, at a – or, the site-specific ground motions vary? 00:52:35.579 --> 00:52:38.420 What are the variances there? 00:52:38.420 --> 00:52:44.250 And try to correlate the ground motion intensities with path-specific properties – 00:52:44.250 --> 00:52:49.610 you know, that would be able to capture, say, basin amplifications and basin edge 00:52:49.610 --> 00:52:54.210 effects and things like that – coupling of the source into the structure. 00:52:54.210 --> 00:52:57.710 And then there’s this idea of spectral correlation. 00:52:57.710 --> 00:53:01.720 How are the adjacent frequencies? How well-correlated are they? 00:53:01.720 --> 00:53:07.560 And that impacts building fragility very, very strongly. 00:53:07.560 --> 00:53:10.700 And there’s evidence that stochastic methods 00:53:10.700 --> 00:53:14.890 don’t produce the right spectral correlation. 00:53:14.890 --> 00:53:18.980 We’ve worked with Jeff Bayless and Norm a bit and provided them 00:53:18.980 --> 00:53:22.400 simulations that they’ve looked at. That’s work in progress, 00:53:22.400 --> 00:53:24.430 but that’s something that the simulations – and there might be 00:53:24.430 --> 00:53:28.390 more to do with the source that has to be modified than with the 00:53:28.390 --> 00:53:31.790 wave propagation and the path. But we need to figure that out. 00:53:31.790 --> 00:53:34.240 So then there’s some other things. 00:53:34.240 --> 00:53:38.950 We’re working specifically right now on, what is the – what are we missing 00:53:38.950 --> 00:53:42.900 by using 500 meter per second as the minimum shear wave speed? 00:53:42.900 --> 00:53:46.599 And we’ve looked at – compared one – you know, 500 meters per second with 00:53:46.599 --> 00:53:52.710 a minimum shear wave speed of 250 and try to quantify what that effect is. 00:53:52.710 --> 00:53:55.820 And, in the same study, we’re also looking at fault geometry. 00:53:55.820 --> 00:53:58.240 How does the fault dip contribute to the asymmetry 00:53:58.240 --> 00:53:59.520 we see across the Hayward Fault? 00:53:59.520 --> 00:54:02.770 Because the fault dips towards the east, towards the low wave speeds in the 00:54:02.770 --> 00:54:06.940 East Bay Hills, and that contributes to the amplifications that we see there. 00:54:06.940 --> 00:54:09.400 We’d like to run suites of ruptures for the same magnitude, 00:54:09.400 --> 00:54:13.150 but the same fault geometry and hypocenter and, really, that’s the study 00:54:13.150 --> 00:54:17.770 that we need to do to generate data for looking at systematic path effects. 00:54:17.770 --> 00:54:20.710 And then there’s a whole universe of moderate-earthquake work that 00:54:20.710 --> 00:54:27.119 we could be doing, trying to understand how well the current model fits the 00:54:27.119 --> 00:54:29.910 observed waveforms for moderate earthquake as a function of path 00:54:29.910 --> 00:54:34.010 and site and frequency, distance. And with that comes all the challenges 00:54:34.010 --> 00:54:36.610 of data and the – and, you know, the challenges of getting the 00:54:36.610 --> 00:54:40.490 source parameters, the location depth, mechanism, moment right. 00:54:40.490 --> 00:54:44.460 There are directivity effects that can be exposed in some of the magnitude, 00:54:44.460 --> 00:54:49.060 say, 4-ish earthquakes that we need to consider, 00:54:49.060 --> 00:54:53.240 or try to find a way to not have that bias the results. 00:54:53.240 --> 00:54:56.280 And, of course, getting all the data from diverse instruments 00:54:56.280 --> 00:54:59.980 and networks is challenging too. So I’ll stop there. 00:54:59.980 --> 00:55:02.750 Thanks for your attention, and I look forward to your questions 00:55:02.750 --> 00:55:05.340 and talking with people today. Thank you. 00:55:05.340 --> 00:55:10.280 [Applause] 00:55:10.280 --> 00:55:12.480 - Great. Thank you. That was very interesting. 00:55:12.480 --> 00:55:14.600 Questions for our speaker? 00:55:15.920 --> 00:55:23.440 [Silence] 00:55:24.240 --> 00:55:26.040 - Hey, Artie. Thanks. That was really impressive, 00:55:26.040 --> 00:55:30.970 and it was great to get a whole tour of everything that’s been going on. 00:55:30.970 --> 00:55:35.240 There’s a lot – I have a lot of questions, but [chuckles], one thing that I’m curious 00:55:35.240 --> 00:55:39.280 about from, you know, throughout your whole talk is that you really focus on 00:55:39.280 --> 00:55:42.750 this being a fully deterministic method. But then you note that you’re 00:55:42.750 --> 00:55:47.410 going down to 9-meter spacing. But that doesn’t exist in the 00:55:47.410 --> 00:55:50.980 velocity models. Question mark? - Yep. It does not. 00:55:50.980 --> 00:55:55.820 - So how does that sort of – so you’re resampling … 00:55:55.820 --> 00:55:58.020 - Yes. - But you don’t have the resolution there. 00:55:58.020 --> 00:55:59.989 So how does that sort of trade off? And can you comment on that? 00:55:59.989 --> 00:56:00.989 - Yeah, so … 00:56:00.989 --> 00:56:03.650 - How is that similar or different than a stochastic-type method? 00:56:03.650 --> 00:56:07.220 - Well, we’re just interpolating the model. 00:56:07.220 --> 00:56:13.760 So when we – you know, we went to run on the Vulcan computer in 2012, 00:56:13.760 --> 00:56:17.420 and that is a – had a big endian/ little endian problem. 00:56:17.420 --> 00:56:23.400 And the [inaudible] VM software and the model couldn’t handle that. 00:56:23.400 --> 00:56:28.240 So Anders basically came up with a different – you know, essentially 00:56:28.240 --> 00:56:31.710 the same hierarchical representation of the model, but we were starting to 00:56:31.710 --> 00:56:35.710 always sub-sample the model. And the model has – you know, 00:56:35.710 --> 00:56:39.940 the USGS model, at the surface, has 100-meter lateral resolution 00:56:39.940 --> 00:56:45.090 and 25 meters in depth. And we were, years ago, sub-sampling that. 00:56:45.090 --> 00:56:48.530 So now, we’re just interpolating. 00:56:48.530 --> 00:56:53.520 So it’s essentially just a linear interpolation. 00:56:54.340 --> 00:56:59.060 That’s how we render the model onto the computational mesh. 00:56:59.069 --> 00:57:03.119 So there’s no – there’s no – there’s no further information that’s put in there. 00:57:03.120 --> 00:57:06.400 It’s just a simple linear interpolation. 00:57:07.820 --> 00:57:12.720 We don’t – I mean, we don’t know 68 billion things about the Earth 00:57:12.720 --> 00:57:16.799 [chuckles] times five, you know, material parameters. 00:57:16.799 --> 00:57:22.940 But that’s what we’re left with trying to do is try to calculate the motions there. 00:57:22.940 --> 00:57:26.860 So the – a lot of the high frequencies come from the source, obviously, 00:57:26.860 --> 00:57:29.790 from those small asperities and the geometry of the 00:57:29.790 --> 00:57:38.480 slip distribution and rake and all that. And short rise time of small asperities. 00:57:38.480 --> 00:57:45.040 So that – I mean, it is a stochastic in that – you know, Rob’s and Arben’s 00:57:45.040 --> 00:57:47.960 slip generator rolls the dice, and it comes up with a random 00:57:47.960 --> 00:57:53.480 distribution and, you know, shapes it to fit along [inaudible] statistics. 00:57:53.480 --> 00:57:54.780 And that’s how the source model – 00:57:54.780 --> 00:57:58.690 so that’s where we get the high frequencies from. 00:57:58.690 --> 00:58:03.340 But in terms of wave propagation, we’re using a very – you know, we’re – 00:58:03.340 --> 00:58:07.640 a representation of the Earth is at much longer wavelengths than the 00:58:07.640 --> 00:58:10.880 shortest wavelength of the simulation. - Right. 00:58:14.020 --> 00:58:17.960 - Another thing we’d like to look at is – Arben’s done a lot of this – 00:58:17.960 --> 00:58:22.560 is introducing stochastic heterogeneity. And then that has more – 00:58:22.560 --> 00:58:25.640 even more knobs to twist of what is the correlation length, 00:58:25.640 --> 00:58:30.010 and how is the – what is the ratio of the horizontal to vertical correlation lengths. 00:58:30.010 --> 00:58:33.520 What is the amplitude? How does that taper with depth? 00:58:33.520 --> 00:58:37.839 And that’s one thing, you know, we’re hoping to be able to look at 00:58:37.840 --> 00:58:43.100 more systematically. And Arben could speak with better authority on that. 00:58:44.860 --> 00:58:52.700 [Silence] 00:58:53.580 --> 00:58:58.950 - You had maps that both look like they had very fine resolution and then, 00:58:58.950 --> 00:59:02.230 towards the end, you had a map that was sort of gridded at, I think, 00:59:02.230 --> 00:59:04.710 closer to your 2 kilometers. - Yeah. 00:59:04.710 --> 00:59:10.820 - Could you explain what your actual output resolution is? 00:59:10.820 --> 00:59:21.180 - Right. So this is a map derived – well, this is a – say, these maps – 00:59:21.180 --> 00:59:27.060 these PGV maps are direct output of SW4 at the resolution of the calculation. 00:59:27.060 --> 00:59:32.010 So, at 12-1/2 meters for the – for the 3D case, 00:59:32.010 --> 00:59:35.220 and I think it’s 20 meters for the 1D case. 00:59:35.220 --> 00:59:38.080 And that just says, you know, as you run the calculation at 00:59:38.080 --> 00:59:44.599 every time step, what is the maximum horizontal X or Y component? 00:59:44.599 --> 00:59:47.790 So that is not orientation-specific. 00:59:47.790 --> 00:59:49.880 But it’s at the resolution of the calculation. 00:59:49.880 --> 00:59:55.910 So you can see – and the movies are made at the resolution of the calculation. 00:59:55.910 --> 00:59:59.730 And then this ratio map is from – I’ve now interpolated the 00:59:59.730 --> 01:00:06.130 1D from 20 meters to 12-1/2 meters and formed this ratio yesterday, and – 01:00:06.130 --> 01:00:13.930 [laughs] – and it’s the topography, again, which is taken from the USGS model – 01:00:13.930 --> 01:00:19.050 from your model and then interpolated to the computational mesh. 01:00:19.050 --> 01:00:24.020 So from 100-meter spacing, I think, down to 12. 01:00:24.920 --> 01:00:29.140 Now, we – but I didn’t show a map, but imagine, every 2 kilometers, 01:00:29.140 --> 01:00:32.170 there’s a seismic station throughout these domains. 01:00:32.170 --> 01:00:35.559 And then that generates SAC files – three-component SAC files – 01:00:35.559 --> 01:00:39.510 you know, three files for every site times – you know, it’s – 01:00:39.510 --> 01:00:44.200 what is it? I think it’s 59 times 39. 01:00:44.200 --> 01:00:51.960 It’s 2,300, you know, sites. And I have three-component 01:00:51.960 --> 01:00:56.950 SAC files that I then measure ground motion intensity measures. 01:00:56.950 --> 01:01:01.220 So PGV, PGA, spectral – RotD50 spectral accelerations. 01:01:01.220 --> 01:01:04.640 Then we have a big pandas table of all that data. 01:01:04.640 --> 01:01:10.150 So then we can form ratios of the 1D to the – or, 3D to the 1D. 01:01:10.150 --> 01:01:14.599 And that’s at this resolution. So this is where I have waveforms. 01:01:14.599 --> 01:01:20.339 So the very fine scale is something that’s output by SW4. 01:01:20.340 --> 01:01:24.259 But these coarser are where I actually have time histories. 01:01:25.720 --> 01:01:29.400 - And getting back to the minimum shear wave speed, I’m curious. 01:01:29.400 --> 01:01:38.900 Have you looked to see, with respect to, say, GMPEs, that if you had included 01:01:38.900 --> 01:01:45.099 the correct Vs30, where do you start to see a greater discrepancy between the – 01:01:45.099 --> 01:01:49.690 or, do you see a greater discrepancy between the simulations and 01:01:49.690 --> 01:01:56.880 what a GMPE would predict as you go to maybe higher frequencies where 01:01:56.880 --> 01:02:01.020 the 500-meter-per-second cutoff is … - Yeah. 01:02:01.029 --> 01:02:04.130 - … no longer valid? - So – good question – so we’ve – 01:02:04.130 --> 01:02:09.760 what we’ve – what I’ve done so far is look at ratios, essentially, 01:02:09.760 --> 01:02:16.760 of the higher-resolved 250-meter-per-second minimum 01:02:16.770 --> 01:02:21.710 shear wave speed to the 500-meter-per- second minimum shear wave speed. 01:02:21.710 --> 01:02:25.780 And then looked at maps at the resolution of the calculation. 01:02:25.780 --> 01:02:30.849 And, you know, it shows that, yeah, where – in the East Bay Flats, 01:02:30.849 --> 01:02:36.150 there can be isolated locations where the – where the motion is – 01:02:36.150 --> 01:02:40.340 the PGV is amplified by as much as, say, a factor of 3. 01:02:40.340 --> 01:02:43.800 But, on average, it’s about 25%. That doesn’t say – that doesn’t 01:02:43.809 --> 01:02:46.890 answer your question about GMPEs, but that’s one thing we’ve done. 01:02:46.890 --> 01:02:50.799 And I can show you time histories that say, yeah, site response matters. 01:02:50.800 --> 01:02:56.320 You know, if we lower the shear wave speed, we get larger ground motions. 01:02:57.940 --> 01:03:03.680 What we’re – work-in-progress now is to look at that more specifically. 01:03:03.690 --> 01:03:08.319 And Arben is going to take our 500-meter-per-second results and 01:03:08.319 --> 01:03:12.400 then apply some corrections to make them – you know, 01:03:12.400 --> 01:03:16.580 that correct them for, say, a 250-meter-per-second minimum 01:03:16.580 --> 01:03:20.700 shear wave speed and then compare those with what we get 01:03:20.700 --> 01:03:25.109 from the calculation to see how some of the empirical methods – 01:03:25.109 --> 01:03:31.060 and I think that’s using a method that he and others at URS developed. 01:03:31.060 --> 01:03:33.790 We haven’t looked at it so much in terms of GMPEs. 01:03:33.790 --> 01:03:36.799 I know I’ve calculated those, but I can’t remember if there’s 01:03:36.799 --> 01:03:38.299 anything dramatic that comes out. 01:03:38.299 --> 01:03:41.010 I mean, it’s a problem in these correlations that the minimum 01:03:41.010 --> 01:03:44.980 shear wave speed only goes – is 500. You see there’s a big cluster of points 01:03:44.980 --> 01:03:51.750 there that, you know, in the real world, would extend to much lower Vs30. 01:03:51.750 --> 01:03:54.720 So we’d like – we’d like to look at that and see – you know, 01:03:54.720 --> 01:03:59.130 it’d be interesting to know that, do the empirical-based corrections – 01:03:59.130 --> 01:04:03.180 you know, how well do they do compared to a full physics calculation? 01:04:03.800 --> 01:04:09.100 And, you know, if – you know, earthquake engineers often would say, 01:04:09.110 --> 01:04:12.200 well, this isn’t a physical parameter. You know, I’m treating it 01:04:12.200 --> 01:04:16.500 like I’m doing a physical calculation. I have a profile at a site. 01:04:16.500 --> 01:04:23.240 I know I can calculate what Vs30 is and what the depth of Z-1.0 and Z-2.5 are. 01:04:23.250 --> 01:04:28.460 But, you know, in the – in the GMPE world, see, these are more 01:04:28.460 --> 01:04:34.980 proxy parameters that are broadly representative of the region or the site. 01:04:34.980 --> 01:04:40.059 And so that’s a slight difference between 01:04:40.060 --> 01:04:43.800 what we’re doing and what’s done in practice. 01:04:46.700 --> 01:04:49.120 - So I think my question is a follow-up on Brad’s. 01:04:49.130 --> 01:04:52.850 And mine – your house and my house are close together in the 01:04:52.850 --> 01:04:55.120 East Bay on the west side of the fault. 01:04:55.120 --> 01:04:58.099 I get earthquake insurance. I pay for earthquake insurance. 01:04:58.099 --> 01:05:01.599 Your results suggest that maybe I don’t need to do that. 01:05:01.600 --> 01:05:03.640 - Yeah. [laughs] [laughter] 01:05:03.640 --> 01:05:06.780 Right. - So I guess my question is, and … 01:05:06.780 --> 01:05:08.140 - It’s something like that. Like that. 01:05:08.141 --> 01:05:12.579 - … you just answered it for the geotechnical properties. 01:05:12.579 --> 01:05:14.880 Have you done enough sensitivity studies – I know 01:05:14.880 --> 01:05:18.200 you’ve looked at the South Napa earthquake and other earthquakes – 01:05:18.200 --> 01:05:26.540 this pattern of lower amplifications or de-amplifications west of the fault, 01:05:26.540 --> 01:05:30.640 that doesn’t – that’s not going to be impacted by change in epicenter 01:05:30.650 --> 01:05:34.600 or magnitude or – should I get insurance? 01:05:34.600 --> 01:05:36.810 Do I still need insurance? [laughter] 01:05:36.810 --> 01:05:39.640 - You’re really going to the bottom line there, right? 01:05:39.640 --> 01:05:46.119 So there’s an important caveat with a map like this, which is, 01:05:46.119 --> 01:05:50.829 you know, don’t mistake this for representing reality. 01:05:50.829 --> 01:05:54.480 Because we haven’t even honored, you know, your best estimate – 01:05:54.480 --> 01:05:57.920 your collective USGS best estimate of the geologic structure 01:05:57.920 --> 01:06:00.980 by using that minimum shear wave speed of 500. 01:06:00.980 --> 01:06:04.340 And that’s – we had to choose something. I couldn’t use 80, 01:06:04.340 --> 01:06:08.980 or else I’d be looking at, you know – you know, a 1-hertz simulation. 01:06:08.980 --> 01:06:16.500 And so we had to make choices there, and that – we’re trying to investigate, 01:06:16.510 --> 01:06:19.430 what are the consequences of that choice. 01:06:19.430 --> 01:06:21.720 And that’s by doing these different calculations. 01:06:21.720 --> 01:06:26.250 And we can do them at 2-1/2 hertz and show that, yeah, we are 01:06:26.250 --> 01:06:32.160 indeed missing amplification. These maps would show less asymmetry 01:06:32.160 --> 01:06:40.420 across the Hayward Fault, but I don’t have an exactly how much right now. 01:06:40.420 --> 01:06:43.930 And that’s work in – you know, that will be, you know, the subject of 01:06:43.930 --> 01:06:49.860 another, you know, study, I think, which is in progress. 01:06:51.420 --> 01:06:56.420 So that's an important caveat, that this shouldn’t necessarily, 01:06:56.430 --> 01:06:59.750 you know, drive you to think – well, I could also – I mean, 01:06:59.750 --> 01:07:03.280 I could be on the – on the ophiolite. [laughs] 01:07:03.280 --> 01:07:06.230 You know, maybe Cal State-Hayward – is that in there? 01:07:06.230 --> 01:07:08.660 Or maybe not, I guess. I think it’s north of there. 01:07:08.660 --> 01:07:13.380 But, you know, is that really as fast as the model? 01:07:13.380 --> 01:07:20.200 And does it go all the way up to the surface with 3.5 kilometers per second? 01:07:20.200 --> 01:07:22.230 You know, we have – there are lots of 01:07:22.230 --> 01:07:25.740 geotechnical measurements in Oakland in the flats. 01:07:25.740 --> 01:07:31.119 And they – we’ve, you know, reported Vs30s from, say, 200 to 400. 01:07:31.120 --> 01:07:34.780 And they’re not being honored in this calculation. 01:07:34.780 --> 01:07:41.240 So, you know, I mean, when I say we want to do the highest frequency most 01:07:41.250 --> 01:07:46.530 realistic model, that’s where, you know, we would like to be efficient. 01:07:46.530 --> 01:07:50.000 And we need even more resources – bigger, more powerful computers, 01:07:50.000 --> 01:07:53.910 more grid points, mesh refinement in the curvilinear grid – 01:07:53.910 --> 01:07:58.420 to be able to honor those geotechnical measurements. 01:08:00.280 --> 01:08:03.720 You know, I usually get a much harder time from earthquake engineers 01:08:03.730 --> 01:08:07.900 and geotechnical engineers [laughs] on that topic. 01:08:07.900 --> 01:08:13.120 - Artie, thank you very much for coming over and giving such a nice talk. 01:08:13.120 --> 01:08:18.580 I’d like you to look into the future. And when should we expect to 01:08:18.580 --> 01:08:23.980 be able to run calculations up to 10 hertz and for velocities of 01:08:23.980 --> 01:08:26.880 100 meters per second at the surface? - Yeah, wow. 01:08:26.880 --> 01:08:28.960 That’s a great question, Tom. 01:08:28.960 --> 01:08:33.670 And I – one thing I can show about this is, you know, past performance. 01:08:33.670 --> 01:08:40.401 So this plot shows, as a function of frequency on the X axis, and number 01:08:40.401 --> 01:08:46.960 of grid points, or total memory of a calculation on the – on the vertical axis. 01:08:46.960 --> 01:08:52.130 And then this solid gray line is the resources that you need to resolve 01:08:52.130 --> 01:08:58.210 a certain frequency for this domain – 180 – or, 120 by 80 by 30 kilometers. 01:08:58.210 --> 01:09:03.430 And then we’ve got a couple of points here, which is – you know, for – in 2012, 01:09:03.430 --> 01:09:07.190 for the California Academy of Sciences, we did – and similar to what we did 01:09:07.190 --> 01:09:12.880 with Brad in the 2010 paper, you know, here’s a 1/2 hertz calculation 01:09:12.880 --> 01:09:15.900 that took a billion grid points. And that was in 2012, 01:09:15.900 --> 01:09:21.239 and we ran that with WPP – an older version – an older code. 01:09:21.239 --> 01:09:24.960 And then, more recently – let’s see. 01:09:24.960 --> 01:09:29.659 So the – to get from the solid gray line is without mesh refinement, but with 01:09:29.659 --> 01:09:33.750 mesh refinement – so that factor of 6. So we dropped down, which means 01:09:33.750 --> 01:09:36.870 we can push further out in the frequency direction. 01:09:36.870 --> 01:09:43.160 And then, you know, there’s the Vulcan calculation in 2014 to 2 hertz. 01:09:43.160 --> 01:09:47.900 And that was, like, 55 billion grid points and was a big calculation at the time. 01:09:47.900 --> 01:09:52.609 And then we started on the ECP in those green – some of those green dots. 01:09:52.609 --> 01:09:56.250 This cluster of dots is what we’ve done in the last few years. 01:09:56.250 --> 01:09:58.270 And then this is the Sierra calculation. 01:09:58.270 --> 01:10:02.300 Now, the total memory of Sierra’s GPUs is up here. 01:10:02.300 --> 01:10:07.850 So we could fit – you know, we could fit this calculation on Sierra. 01:10:07.850 --> 01:10:12.380 But we’re not – we might not – we might need it for several days. 01:10:12.380 --> 01:10:16.080 You know, so that’s a challenge. 01:10:16.920 --> 01:10:19.480 Let’s see. And one more point here is that, like, 01:10:19.480 --> 01:10:25.860 we realized the Cori calculation we did in June – you know, we’re below – 01:10:25.860 --> 01:10:29.530 fairly far below Cori’s total memory, but we’re – you know, we’re not 01:10:29.530 --> 01:10:32.430 going to get the machine for a week to run – you know, and what we’d 01:10:32.430 --> 01:10:36.139 like to do are the red dots, which is the 10-hertz calculation. 01:10:36.140 --> 01:10:39.520 But I’m not answering your question, still, though. 01:10:39.520 --> 01:10:41.100 [laughter] 01:10:41.100 --> 01:10:46.420 So, you know, it took us – you know, I don’t know, what, six years. 01:10:46.420 --> 01:10:51.440 And I don’t think we were as careful doing things in the past, you know, 01:10:51.440 --> 01:10:55.220 as we are now, thanks to the work that Anders and Bjorn and 01:10:55.220 --> 01:10:58.480 Hans Johansen and others had done to make the code. 01:10:58.489 --> 01:11:02.810 And, of course, now the GPU version – you know, and I’ve drank the 01:11:02.810 --> 01:11:08.440 Kool-Aid on – you know, on the GPUs. It’s really the way to go. 01:11:08.440 --> 01:11:12.960 Because we were running on a small fraction of the Sierra machine, 01:11:12.960 --> 01:11:16.240 and we were able to reproduce what we did on all of Cori. 01:11:16.240 --> 01:11:21.830 And we have a much better chance of getting a small part of Sierra or its – you 01:11:21.830 --> 01:11:27.630 know, a computer that’s going to be very similar to Cori that’ll be available to us. 01:11:27.630 --> 01:11:33.210 So we think that we can get to 10 hertz with the existing architecture. 01:11:33.210 --> 01:11:37.080 We think we can probably do it with – we’re hoping to do it in the next 01:11:37.080 --> 01:11:41.100 few months on Sierra before it goes behind the fence. 01:11:41.100 --> 01:11:45.540 And – but in turn – you know, so that’s what you can do 01:11:45.540 --> 01:11:50.040 with something in the top 10 of most powerful computers, right? 01:11:50.040 --> 01:11:56.740 But, you know, it might be a few years, you know, before this is now available 01:11:56.750 --> 01:12:02.110 on sort of commodity, you know, GPU clusters that a professor 01:12:02.110 --> 01:12:07.159 could buy with their start-up money or organizations could purchase. 01:12:07.159 --> 01:12:14.420 And, you know, it’ll be really exciting to see how we might use these fully 3D, 01:12:14.420 --> 01:12:20.660 you know, physics-based simulations to look at hazard and risk. 01:12:20.660 --> 01:12:23.510 And that’s kind of what we’re trying to do with this – 01:12:23.510 --> 01:12:25.580 with this project, you know. 01:12:25.580 --> 01:12:28.320 And there’s a whole engineering part of this that you haven’t 01:12:28.320 --> 01:12:32.850 heard about that Dave could come back and talk about. 01:12:32.850 --> 01:12:35.940 But it might not be that far away, you know, if things keep going. 01:12:35.940 --> 01:12:41.540 And this ECP – part of the point is to try to figure out, you know, 01:12:41.540 --> 01:12:44.380 what happens when you’re running on half a million cores. 01:12:44.380 --> 01:12:46.490 You know, or all these GPUs. 01:12:46.490 --> 01:12:49.940 And, you know, how do – I mean, a lot has to go right for these 01:12:49.940 --> 01:12:52.350 calculations to work [chuckles], obviously. 01:12:52.350 --> 01:12:59.969 And then that is really a – you know, the undiscovered country that 01:12:59.969 --> 01:13:03.250 we’re going into is to figure out, how do we overcome 01:13:03.250 --> 01:13:04.910 these unanticipated challenges? 01:13:04.910 --> 01:13:11.140 Like, all of this work, up until Sierra, was CPU-based, MPI communications. 01:13:11.140 --> 01:13:13.920 You know, it’s a lot – it’s not that different, really. 01:13:13.929 --> 01:13:17.690 There’s some improvement going from, you know, improving the inter-node – 01:13:17.690 --> 01:13:21.770 when you have 68 cores in one node, you can use fast shared memory. 01:13:21.770 --> 01:13:23.320 And then, when you have to go in between nodes, 01:13:23.320 --> 01:13:25.270 you have to use MPI, and that’s what they did. 01:13:25.270 --> 01:13:31.610 And they showed they got a factor of 3 improvement in performance. 01:13:31.610 --> 01:13:35.360 So it might not be that far. 01:13:35.360 --> 01:13:40.040 Future grad students could be – their thesis could be on, oh, I ran a catalog of, 01:13:40.040 --> 01:13:45.000 you know, Hayward Fault simulations – 10,000 events resolved to 5 hertz. 01:13:45.740 --> 01:13:47.840 Pretty exciting. 01:13:49.320 --> 01:13:57.360 [Silence] 01:13:58.160 --> 01:14:02.400 - Artie, just a quick question. Your Bay Area 3D velocity model, 01:14:02.400 --> 01:14:08.219 did it include other faults? Or the 3D simulation of other faults, 01:14:08.219 --> 01:14:14.900 like the Rodgers Creek, and how that affects propagation? Or the Calaveras 01:14:14.900 --> 01:14:19.520 or other faults off the Hayward? - Yeah. 01:14:19.520 --> 01:14:25.740 So, I mean, it’s – the model that was built in this office – it’s that same model 01:14:25.740 --> 01:14:30.000 that, you know, we can download off of your – the USGS website. 01:14:30.000 --> 01:14:33.060 So it has the 3D fault geometry in it. 01:14:33.060 --> 01:14:37.860 And you can’t – you can’t see it so much – maybe in the animations. 01:14:37.860 --> 01:14:43.040 You can see that crossing faults impacts ground motion. 01:14:44.220 --> 01:14:48.000 You know, you can see, as waves hits – I think, like, 01:14:48.010 --> 01:14:52.910 the Concord Fault or Green Valley – sometimes – like, this structure 01:14:52.910 --> 01:14:54.250 right up here is going to show up. 01:14:54.250 --> 01:14:57.680 Well, you can see it down here in the Calaveras. 01:14:57.680 --> 01:15:00.780 So you can see that, and it’s the structure, and there’s material 01:15:00.780 --> 01:15:05.800 heterogeneity across those faults that indeed impacts ground motion. 01:15:06.860 --> 01:15:10.880 And, you know, in terms of capabilities, I mean, there’s nothing preventing us 01:15:10.880 --> 01:15:15.500 from running a simulation of a Calaveras or a Napa or a Rodgers Creek. 01:15:15.500 --> 01:15:18.730 I mean, we did that, you know, with Brad a few years ago. 01:15:18.730 --> 01:15:22.840 There were other scenarios. It’s just, you know, a matter of 01:15:22.840 --> 01:15:24.800 choosing what scenarios you want to look at. 01:15:24.800 --> 01:15:28.360 I’ve look at the Greenville Fault for Livermore because that’s – you know, 01:15:28.360 --> 01:15:31.980 drives hazard at Lawrence Livermore Lab out here. 01:15:31.980 --> 01:15:37.110 So I’m not sure if I’m answering your question or not, but it is – it is possible 01:15:37.110 --> 01:15:44.940 to run other simulations in other faults. And those faults show up in the model 01:15:44.940 --> 01:15:53.380 in terms of wave propagation effects. I mean, we saw – in fact, the – well, 01:15:53.380 --> 01:15:57.910 you can see the San Andreas, in fact, enters into, you know, 01:15:57.910 --> 01:16:00.640 the – showing up as wave propagation effects. 01:16:01.560 --> 01:16:04.060 I’m sorry. I’m not sure if I’m answering your question. 01:16:04.060 --> 01:16:06.340 Let me know if I’m not. 01:16:07.820 --> 01:16:09.820 - Final questions for Artie? 01:16:10.780 --> 01:16:13.560 All right. Well, thanks again, and thanks, everyone, for coming. 01:16:13.570 --> 01:16:17.280 Our next Wednesday seminar will be at 1:30 in the afternoon, as a reminder. 01:16:17.280 --> 01:16:21.000 And then tomorrow at 2:00, you’re welcome to join us for coffee and more 01:16:21.000 --> 01:16:24.780 discussion on tall building response to earthquakes in San Francisco. 01:16:24.780 --> 01:16:26.540 Thanks. - All right. Thank you. 01:16:26.540 --> 01:16:30.780 [Applause] 01:16:30.780 --> 01:16:36.140 - There’s still one more spot [inaudible] this afternoon is anyone [inaudible]. 01:16:36.140 --> 01:16:38.340 - Oh, great.