WEBVTT 1 00:00:00.100 --> 00:00:01.549 STFC-RAL-CR03 R61: Okay, Congressman. 2 00:00:01.870 --> 00:00:05.860 STFC-RAL-CR03 R61: Hello, everyone, and welcome to today's seminar. 3 00:00:06.180 --> 00:00:12.809 STFC-RAL-CR03 R61: Let me introduce our speaker. I… I will read, the bio, and it's, 4 00:00:12.980 --> 00:00:19.670 STFC-RAL-CR03 R61: Yeah, I just could, like, an equation, I couldn't skip anything, or like a theorem, it's difficult. 5 00:00:20.090 --> 00:00:26.189 STFC-RAL-CR03 R61: So… Our, guest speaker today, Iran, is, 6 00:00:26.750 --> 00:00:35.139 STFC-RAL-CR03 R61: Currently a theme lead in the computational mathematics in SDS's Scientific Computing Department. 7 00:00:35.240 --> 00:00:40.780 STFC-RAL-CR03 R61: Where he oversees books in numerical linear algebra. 8 00:00:41.110 --> 00:00:46.669 STFC-RAL-CR03 R61: And inverse problems, and leads a community of practice. 9 00:00:47.610 --> 00:00:50.350 STFC-RAL-CR03 R61: in medical algorithms. 10 00:00:50.580 --> 00:00:56.999 STFC-RAL-CR03 R61: He completed his… PhD at the University of Oxford in 2010. 11 00:00:57.190 --> 00:01:03.449 STFC-RAL-CR03 R61: On preconditioning iterative methods called partial differential equations. 12 00:01:03.760 --> 00:01:06.690 STFC-RAL-CR03 R61: Constraint optimization. 13 00:01:08.560 --> 00:01:22.700 STFC-RAL-CR03 R61: for constraint optimization. Prior to joining SCFC, he spent 2 years as a postdoctoral fellow at the University of British Columbia in Vancouver, Canada. 14 00:01:23.140 --> 00:01:32.779 STFC-RAL-CR03 R61: His research interests lie in the intersection of numerical linear algebra, the numerical solution of EDs. 15 00:01:32.950 --> 00:01:43.240 STFC-RAL-CR03 R61: and continuous optimization. He's particularly interested in the use of iterative methods to solve large, fast systems of equations. 16 00:01:43.300 --> 00:01:58.389 STFC-RAL-CR03 R61: and the acceleration of such methods using preconditioners. As well as theoretical issues surrounding these methods, he's also interested in using them to efficiently solve problems from applications. 17 00:01:58.540 --> 00:02:05.230 STFC-RAL-CR03 R61: With a particular interest in the systems arising in the optimal controller. 18 00:02:06.100 --> 00:02:12.580 STFC-RAL-CR03 R61: He has overseen the development of the internationally renowned HSL, 19 00:02:12.830 --> 00:02:20.260 STFC-RAL-CR03 R61: which, H stands for FireWell, a mathematical Software Library since 2019. 20 00:02:20.520 --> 00:02:36.349 STFC-RAL-CR03 R61: In optimization, he's interested in the development of novel methods for solving nonlinear least squares problems, and he's the primary developer of the Ralph Fitz software, which is now part of AMPs. 21 00:02:36.970 --> 00:02:42.630 STFC-RAL-CR03 R61: the MD, company, data analytics library. 22 00:02:43.310 --> 00:02:48.639 STFC-RAL-CR03 R61: So, love to hear more from you. So welcome to Zurong. 23 00:02:51.600 --> 00:03:03.870 STFC-RAL-CR03 R61: Thank you very much, and, yeah, thank you very much for the opportunity to come and speak to you. So, yeah, I don't think we don't do enough of cross-pollination within SDFC, so this is a wonderful opportunity. 24 00:03:04.700 --> 00:03:13.709 STFC-RAL-CR03 R61: There's more, but I'm going to do a very high-level touch, I'm not going to delve into any details of all things that we do, there's lots more there. 25 00:03:13.710 --> 00:03:30.450 STFC-RAL-CR03 R61: There's lots of QR codes around if you want to, scan these and share the slides afterwards if you want to go to the pieces. Here's the website, the theme numerical.rl.ac.uk. You can find more old papers and details of all the projects we're involved in in Goldfinder. 26 00:03:30.540 --> 00:03:33.800 STFC-RAL-CR03 R61: More information on anything, there. 27 00:03:34.550 --> 00:03:37.020 STFC-RAL-CR03 R61: So, to start with, I just wanted to give a… 28 00:03:37.640 --> 00:03:57.080 STFC-RAL-CR03 R61: sketch a… step back from the competition maths team and see… show how scientific computing is organized, because, again, all these different areas of STFC are all organized very different, being from the kind of opaque looking from the outside in, and from my side looking elsewhere. So, in scientific computing, we have… we're split into 29 00:03:57.570 --> 00:04:00.989 STFC-RAL-CR03 R61: So we split into themes. 30 00:04:01.200 --> 00:04:14.220 STFC-RAL-CR03 R61: And, and programs. So, most of the people in the department live in themes, and so themes, you can kind of think of as a small division. We used to have four divisions. That was broken into, the 31 00:04:14.460 --> 00:04:20.600 STFC-RAL-CR03 R61: Also, themes that we have now, which are a lot more… automatically fixed, and so… 32 00:04:20.750 --> 00:04:24.199 STFC-RAL-CR03 R61: I look at the computational matsey, 33 00:04:24.400 --> 00:04:37.309 STFC-RAL-CR03 R61: We sit along here. They're also a theme for AI for Science, Competition Biology, computational engineering, open science, software engineering platforms and services, and all these, this is where the majority of the people within the bounds sit. 34 00:04:37.700 --> 00:04:42.340 STFC-RAL-CR03 R61: As well as the themes, we have, the programs that are run, in the department. 35 00:04:42.570 --> 00:05:01.660 STFC-RAL-CR03 R61: IRIS, which I guess most people, you know, interact with IRIS a lot. Clsec and the, the, Physical Science Data Infrastructure, these are how money comes into the department, so these are sources of funding. I'll talk a bit about ELC and CoreSec. 36 00:05:01.660 --> 00:05:02.760 STFC-RAL-CR03 R61: So, projects… 37 00:05:02.760 --> 00:05:07.769 STFC-RAL-CR03 R61: Run along programs, and then the people doing the work on the projects live in the themes, and you can… 38 00:05:07.770 --> 00:05:19.960 STFC-RAL-CR03 R61: For a project, you might be a mathematician, plus someone who knows something about AI for science, plus a software engineer, and get them all together to get it to deliver it. 39 00:05:20.770 --> 00:05:25.660 STFC-RAL-CR03 R61: The first one of these programs are going to talk about is ELogic Center, ELC. 40 00:05:25.790 --> 00:05:33.300 STFC-RAL-CR03 R61: So here's the page from the website that explains what it does a little bit. I'm gonna get the QR code to the website if you want to know more. 41 00:05:33.640 --> 00:05:42.839 STFC-RAL-CR03 R61: The Adelphia center provides scientific computing expertise and innovation to help researchers using the SDFC National Labs to do bigger and better science. 42 00:05:43.250 --> 00:05:50.679 STFC-RAL-CR03 R61: So the big picture behind the ALC is the, if you think of ISIS or secularities, it'll be a violent. 43 00:05:51.000 --> 00:06:02.640 STFC-RAL-CR03 R61: These all do different physics and do different science, but essentially, they meet the same infrastructure and the data analysis from a mathematician's point of view. 44 00:06:03.360 --> 00:06:04.689 STFC-RAL-CR03 R61: They… 45 00:06:05.300 --> 00:06:18.230 STFC-RAL-CR03 R61: generally fire stuff at something, look at the stuff that comes in, do some data fitting at the end, so the algorithms and data are impalitable. What used to happen, and still happens sometimes, is that these 46 00:06:18.710 --> 00:06:38.069 STFC-RAL-CR03 R61: the facilities, and possibly even an instrument within the facility, will build that whole infrastructure thing by themselves, and you can… if you can connect dots and help people along and say, if you talk to this person there, they've already solved the same problem you're having now, a few years ago, you can… you can learn from that, or… or you can learn people from CLF can use the same 47 00:06:38.150 --> 00:06:46.270 STFC-RAL-CR03 R61: techniques that, people like ISIS are using and don't use, and just share that knowledge a bit more, help collaboration a bit better. So, this is what 48 00:06:46.600 --> 00:06:52.579 STFC-RAL-CR03 R61: Centrally funded from SDSC Leaderly, Eido Limbly center helps connect those… those dots. 49 00:06:53.240 --> 00:06:56.410 STFC-RAL-CR03 R61: And, lots of our funding, comes from there. 50 00:06:57.220 --> 00:06:58.480 STFC-RAL-CR03 R61: Yeah, that… 51 00:06:58.880 --> 00:07:14.820 STFC-RAL-CR03 R61: thing I'll talk about is COSEC. Cosec stands for, Computational Science Center for Research Communities. So whereas the Logday center is kind of inward-looking and it helps our facilities, COSEC is more output-looking. It sees what are the research communities that the software 52 00:07:14.880 --> 00:07:20.989 STFC-RAL-CR03 R61: the research communities within the UK have developed software. How could we support them as a national lab? 53 00:07:21.190 --> 00:07:30.330 STFC-RAL-CR03 R61: So, originally, it was an EPSRC, we work with EPSC and its EPSC… EPSRC communities, in… 54 00:07:30.580 --> 00:07:42.010 STFC-RAL-CR03 R61: recent years, that's expanded with the EPA, the Research Infrastructure Fund, and so we, you know, reach out to other, communities that don't fit within the EPSRC need. 55 00:07:43.070 --> 00:07:59.179 STFC-RAL-CR03 R61: Within these communities, the ones that we are most involved in is, CCP DCM, so Data Centered, Continuum Continuum Mechanics, so it's developed by an abandonment software, Fiverrid in Phoenix, so we provide some support for them. 56 00:07:59.230 --> 00:08:18.909 STFC-RAL-CR03 R61: Ccbqc, and so CCP and, CCP is Computational Collaborative Project. So helping quantum computing, all the different communities, how can they, access and make use of quantum computing dozens. Ccpi, which is, 57 00:08:19.050 --> 00:08:36.779 STFC-RAL-CR03 R61: Tomography, computational tomography, UKNR in American Relativity, and there's a, CCP in theoretical experimental particle physics, and I'll talk a little bit more, about that later. So, the idea behind these, 58 00:08:37.309 --> 00:08:53.639 STFC-RAL-CR03 R61: COSEC of the program, funds collaboration money, so they can run workshops and get together as a community in universities from the UK to build around there. It also funds some development time, so the staff time within SDFC is kind of like if, you know, research code 59 00:08:53.640 --> 00:09:03.239 STFC-RAL-CR03 R61: you have a… you get a postdoc, and the postdoc does some amazing stuff, some code, it does really good things. The postdoc believes no one really knows what's going on with the code. 60 00:09:03.240 --> 00:09:15.780 STFC-RAL-CR03 R61: And then, the project needs. So it's, course efforts funded, postdocs that don't quit, but STSC will provide the support. If someone leaves, someone else will come in and pick up the 61 00:09:15.920 --> 00:09:21.190 STFC-RAL-CR03 R61: Some of these communities have been going for 60 years, so it's long to a small amount of funding, very constant. 62 00:09:24.260 --> 00:09:35.439 STFC-RAL-CR03 R61: So when does host computational math fit in the UK ecosystem? So, on, this side here, my pointer, we have these 63 00:09:35.440 --> 00:09:54.149 STFC-RAL-CR03 R61: list of things that go in here will change from time to time, but it's essentially the stuff that SDAC says we should be interested in. So, here are the SDAC facilities, and Diamond, maybe. There's the Corset communities we just talked about, the other themes of Inside of computing, other UK labs. 64 00:09:54.290 --> 00:10:05.920 STFC-RAL-CR03 R61: then work with, UKDA and Calum and other places, where we have skills that they need. And, the ISE priority areas, there's a lot to talk about that, and so anything… 65 00:10:06.060 --> 00:10:09.049 STFC-RAL-CR03 R61: People that are working in these areas. 66 00:10:09.230 --> 00:10:16.319 STFC-RAL-CR03 R61: will have a need for mathematical algorithms. They'll have problems they want to solve, mathematically on a computer, and generally they'll want 67 00:10:16.930 --> 00:10:24.080 STFC-RAL-CR03 R61: faster? Can I solve this problem faster, or can I solve it more accurately? And that's where they come to us. 68 00:10:24.330 --> 00:10:32.670 STFC-RAL-CR03 R61: We… we do research, we prove theorems, we show things work, we publish papers, 69 00:10:32.860 --> 00:10:35.620 STFC-RAL-CR03 R61: So we can do research based on these. 70 00:10:36.150 --> 00:10:38.000 STFC-RAL-CR03 R61: Questions that people come to us with. 71 00:10:39.060 --> 00:10:52.560 STFC-RAL-CR03 R61: our work doesn't end with research, it doesn't end with a paper. If it did, we'd better sit at the university, so we write software, high-quality software, as a result of the research, which then gets fed back into the communities. 72 00:10:53.620 --> 00:10:57.479 STFC-RAL-CR03 R61: Sometimes there's no research to be done there, well… 73 00:10:57.760 --> 00:11:09.140 STFC-RAL-CR03 R61: But it's… there's something in the literature, so we kind of know the math aspects, computational math literature, and sometimes it's just like, oh yeah, you need to read this paper where John Brons did this, and… 74 00:11:09.160 --> 00:11:24.859 STFC-RAL-CR03 R61: 2010, you can implement this, and so sometimes it's just to point people to the area in which they might be experts in particle physics, maybe, but they wouldn't know the math. I don't know the particle physics literature, it's reasonable, they don't know, so we can open there. 75 00:11:25.240 --> 00:11:29.519 STFC-RAL-CR03 R61: Sometimes there's research to be done which, isn't within our 76 00:11:29.590 --> 00:11:41.749 STFC-RAL-CR03 R61: We don't know everything, we have some expertise in some areas, but there are other areas we don't have expertise in. But we do go around and get well connected within the world of computational maths community, so we can 77 00:11:41.750 --> 00:11:50.000 STFC-RAL-CR03 R61: People up with researchers at universities elsewhere, and say, this is the person we need to work with, and that's how we can solve your problem. 78 00:11:50.800 --> 00:11:53.620 STFC-RAL-CR03 R61: And we're well connected with, 79 00:11:53.870 --> 00:12:00.459 STFC-RAL-CR03 R61: learned society did maths, so we have three SAM fellows, the Society for Industrial and Applied Maths. 80 00:12:01.040 --> 00:12:17.269 STFC-RAL-CR03 R61: We have a number of IMA, Institute of Maths Applications tablets. We have a member of the Academy for Mathematical Sciences within the theme. So, where there are needs for people to do research questions, or to get funding that solves the questions that people in the 81 00:12:17.300 --> 00:12:29.459 STFC-RAL-CR03 R61: goodies that are maths adjacent have, we can book it for them in these areas and, yeah, make sure there's talks at conferences, in maths conferences that shine a light on the areas that people 82 00:12:29.500 --> 00:12:32.819 STFC-RAL-CR03 R61: In these communities that need help with and so on. 83 00:12:34.830 --> 00:12:38.399 STFC-RAL-CR03 R61: So here's… here's the thief, so… 84 00:12:38.820 --> 00:12:43.910 STFC-RAL-CR03 R61: only 3 groups. There's a group in numerical and algebra. 85 00:12:44.480 --> 00:12:55.259 STFC-RAL-CR03 R61: I group in continuous optimization by Terry Fox, and then I group in ingress problems by Margaret efforts. No, there's any Inverse Problems questions, close enough. 86 00:12:55.810 --> 00:12:59.239 STFC-RAL-CR03 R61: We'll go from… yes, we have a number of… 87 00:12:59.350 --> 00:13:16.619 STFC-RAL-CR03 R61: Yeah, all the way from IM Fellows, that we have, Jennifer and Nick are both, and HIM fellows, and down to a lot of, PhD students that come through, so we have the wide range of, research experience. 88 00:13:16.620 --> 00:13:23.680 STFC-RAL-CR03 R61: And they're a mixed between, people with PhDs in mathematics and software engineers, so there's a mix of… 89 00:13:26.300 --> 00:13:45.660 STFC-RAL-CR03 R61: And, yeah, lots of books. I've discovered, lots of Brutus books, but lots of books in various areas, and I'll click on some of these that we have, yeah, so Jennifer's written algorithms, sparse Linear Systems, Andy Wharton, who's written a book on fined elements, and Nicole's written a book on, trust-reading methods. 90 00:13:45.730 --> 00:13:57.319 STFC-RAL-CR03 R61: John Reed, who used to lead for… Reed's compound before me, has written the book on modern forefront, who's been the full track, committee for many years. 91 00:13:58.680 --> 00:14:03.689 STFC-RAL-CR03 R61: And we do, yeah, lots of other stuff. Software is… high-quality software is… 92 00:14:04.160 --> 00:14:12.480 STFC-RAL-CR03 R61: the core of what we do. If you look at the journal ECM Transactions and mathematical Software, we're the third most, 93 00:14:13.750 --> 00:14:23.149 STFC-RAL-CR03 R61: well, relevant for maths and math, which removes the computational maths, theme, is that it is the fourth most, appeared, published 94 00:14:23.320 --> 00:14:28.670 STFC-RAL-CR03 R61: body in there behind Argonne, Sandia, and University of Texas as Foston. 95 00:14:31.040 --> 00:14:33.960 STFC-RAL-CR03 R61: So, when people… 96 00:14:34.150 --> 00:14:44.850 STFC-RAL-CR03 R61: That's what we do. I like to describe it as we build LEGO bricks. I'm possibly not pouring LEGO bricks like this, but we're looking at Lego bricks like this, so if you… 97 00:14:44.850 --> 00:14:56.010 STFC-RAL-CR03 R61: My kids are slightly older now, but at some point, young kids that like playing label, the beauty of label is from one carefully crafted label brick. I don't really need that guy there. 98 00:14:56.630 --> 00:15:13.410 STFC-RAL-CR03 R61: you can build all sorts of label sets. So, the cat's nose might be in the engine of a Formula 1 car, might be in the cockpit of the Merlinian Falcon that's on the same label bricks. You can plug into different things, and that's how I see the algorithms we design. It's kind of like that. 99 00:15:13.580 --> 00:15:26.280 STFC-RAL-CR03 R61: might have the Formula One car in mind to start with, but then the tool that we build would be able to be the system solver, our optimization solver, and we put in all sorts of different applications, and that kind of thing, I think. 100 00:15:26.800 --> 00:15:31.479 STFC-RAL-CR03 R61: Make things happen that we didn't before it was possible that we chose it. 101 00:15:32.470 --> 00:15:33.349 STFC-RAL-CR03 R61: I don't know. 102 00:15:33.450 --> 00:15:37.849 STFC-RAL-CR03 R61: Shine a light on some of those areas where this has happened, as we go through. 103 00:15:38.360 --> 00:15:48.860 STFC-RAL-CR03 R61: So, I'm going to focus on three different groups. Again, I can't talk about everything, and I'm not going to go into detail of what we do in the groups, but I'm going to try and give a flavor of the stuff we do. 104 00:15:49.580 --> 00:15:51.100 STFC-RAL-CR03 R61: Hopefully, that's helpful. 105 00:15:51.710 --> 00:15:53.920 STFC-RAL-CR03 R61: So, start with numerical linear. 106 00:15:54.500 --> 00:15:59.829 STFC-RAL-CR03 R61: And in particular, solving AX equals B, so linear system equations. 107 00:16:01.340 --> 00:16:07.780 STFC-RAL-CR03 R61: And, think of it as a matrix, A, times an unknown vector, X. 108 00:16:07.890 --> 00:16:11.870 STFC-RAL-CR03 R61: Right-hand side would be, nice square matrix. 109 00:16:12.090 --> 00:16:20.840 STFC-RAL-CR03 R61: And think of n in the order of millions is what we solve. So, if you have a linear system, like this to solve. 110 00:16:21.970 --> 00:16:36.559 STFC-RAL-CR03 R61: Here's Poisson's equation, so a simple PD disguise the PD using finer elements. If you use the method using Gaussian elimination period schools, doing A-levels within Gaussian elimination, solve it on a computer, this is how 111 00:16:36.680 --> 00:16:42.660 STFC-RAL-CR03 R61: As you increase the number of the solution type goes up, skillset for the N cubed, 112 00:16:43.130 --> 00:16:45.199 STFC-RAL-CR03 R61: scale it with this algorithm. 113 00:16:46.880 --> 00:16:58.619 STFC-RAL-CR03 R61: What I haven't told you for this system, so I have to know for 2D Poisson, and for all the problems we're interested in, we're interested in sparse linear systems, so most of the entries are zeros. 114 00:16:58.680 --> 00:17:07.749 STFC-RAL-CR03 R61: So, the number of non-zeros is of the order of the number of entries in a row of the matrix. If you do that. 115 00:17:07.930 --> 00:17:17.669 STFC-RAL-CR03 R61: the graph I showed you last explicitly counted these zeros, that's what you would do if you were in school. You can… you can crop the zeros, hold this as a, 116 00:17:18.040 --> 00:17:19.689 STFC-RAL-CR03 R61: The data structure. 117 00:17:19.810 --> 00:17:39.720 STFC-RAL-CR03 R61: And if you squeeze this a little bit, then there's a correspondence between a sparse matrix and a graph. So the graph will show connectedness between surfaces, entries here. There's a connection between, node 1 and node 4 in the graph. 4 and 2 are connected, it shows there. 118 00:17:39.790 --> 00:17:48.949 STFC-RAL-CR03 R61: And once you have this correspond to your sparse matrix and a graph, then you can use all sorts of cool algorithms from graph theory to be able to use, to do… 119 00:17:49.260 --> 00:18:02.220 STFC-RAL-CR03 R61: Gaussian emissions on a graph, which makes much more powerful, methods for doing this. So, here's my… my base, not taking into account the structure of the migrants. You have this order en cubed, behavior. 120 00:18:03.620 --> 00:18:14.809 STFC-RAL-CR03 R61: Here's the same problem using a sparse direct method, and as the number of unknowns grows, you can solve much, much larger problems with shooting off. 121 00:18:15.120 --> 00:18:16.489 STFC-RAL-CR03 R61: You know, much more. 122 00:18:18.000 --> 00:18:21.290 STFC-RAL-CR03 R61: not quite linearer, mostly, in this case. 123 00:18:21.910 --> 00:18:30.560 STFC-RAL-CR03 R61: And, this is… this is a fundamental idea behind the most popular calls in the HSI world. 124 00:18:30.730 --> 00:18:35.829 STFC-RAL-CR03 R61: HSL has a… has a long history. Started… 125 00:18:35.930 --> 00:18:49.409 STFC-RAL-CR03 R61: back in the 1960s, so this guy, Matt Powell, who, back in… we were the Harwell Lab, or the group, I wasn't there, but the group was part of the Harwell Lab in those days, and they were, they were building… 126 00:18:49.530 --> 00:18:52.740 STFC-RAL-CR03 R61: various, similar to ELC now, different 127 00:18:53.080 --> 00:19:07.339 STFC-RAL-CR03 R61: groups in that area have their own subroutines for solving different problems. Matt Powell came along, he's kind of one of the first research software engineers, and said, let's collect all these, make sure they're good quality quotes, and call it, HSL, HSL was 128 00:19:07.340 --> 00:19:13.610 STFC-RAL-CR03 R61: stood for the Harold Subroutine Library. And so everyone has a canonical version of all these quotes. 129 00:19:13.990 --> 00:19:18.090 STFC-RAL-CR03 R61: So, various people, the bold ones, the people who make big contributions. 130 00:19:18.190 --> 00:19:23.599 STFC-RAL-CR03 R61: in 1990, and how all that was closed down, things were privatized, the group 131 00:19:23.720 --> 00:19:30.109 STFC-RAL-CR03 R61: moved over, competition maths, moved to the Rutherford Lab, and, that's continued since then. 132 00:19:30.930 --> 00:19:43.880 STFC-RAL-CR03 R61: We probably still do have some code from them that it succeeds, but most of the code is written, later. Stuff is still around, but stuff gets added all the time. I joined in 2012, and I've been looking after HSL since 133 00:19:43.990 --> 00:19:45.739 STFC-RAL-CR03 R61: Focusing on the team. 134 00:19:47.350 --> 00:19:48.440 STFC-RAL-CR03 R61: we have… 135 00:19:48.580 --> 00:19:57.899 STFC-RAL-CR03 R61: academic users from all around the world. So if you're an employer, or a student at a university, or a non-profit research lab. 136 00:19:58.230 --> 00:20:11.850 STFC-RAL-CR03 R61: there's a free license available, so every year we give around 4,000 licenses, for HSL software, and here's some of the, packages that kind of rely on HSL to enable… 137 00:20:12.600 --> 00:20:13.390 STFC-RAL-CR03 R61: Yep. 138 00:20:14.800 --> 00:20:21.580 STFC-RAL-CR03 R61: Lots of, lots of user software, and then all the other stuff that's built on top of that. 139 00:20:22.710 --> 00:20:30.689 STFC-RAL-CR03 R61: We sell licenses commercially, so we offer single user licenses, site licenses, and corporation licenses. 140 00:20:31.010 --> 00:20:33.269 STFC-RAL-CR03 R61: Any print we get is invested 141 00:20:33.520 --> 00:20:46.360 STFC-RAL-CR03 R61: And again, for the LEGO we built a smart linear solver, it's the same solvers used by the Department of Energy, Exxon, 142 00:20:46.440 --> 00:20:54.569 STFC-RAL-CR03 R61: all the Formula 1 teams have a license to optimize, re-extract, insulin, xylinder, water for them. 143 00:20:55.210 --> 00:21:01.270 STFC-RAL-CR03 R61: solution company, and General Electric, so there's not much in common between all these, 144 00:21:01.650 --> 00:21:04.719 STFC-RAL-CR03 R61: Except we all need to solve spast linear systems underneath it. 145 00:21:04.990 --> 00:21:08.460 STFC-RAL-CR03 R61: And that's what they paid us to do that. 146 00:21:09.850 --> 00:21:11.790 STFC-RAL-CR03 R61: Better said that? 147 00:21:12.140 --> 00:21:16.860 STFC-RAL-CR03 R61: Massive amount of software that does this, but if you know something, know your problem. 148 00:21:18.160 --> 00:21:34.870 STFC-RAL-CR03 R61: I've highlighted sparsity, but there's other properties that our matrix can have. If it's symmetric, if it's, if it's unsymmetric, if it's almost symmetric, if it's positive definite, if it's positive definite, but sort of all these different cases, there's a different algorithm you can use. 149 00:21:36.510 --> 00:21:40.789 STFC-RAL-CR03 R61: If you're not sure, talk to us, and we can… we can help. 150 00:21:41.960 --> 00:21:54.900 STFC-RAL-CR03 R61: As well as HSL, we have an open source solver, SIDS, which does the same thing, solves fastening system equations. This is GPU-enabled, so you can use a GPU to solve your linear systems. 151 00:21:55.060 --> 00:22:11.669 STFC-RAL-CR03 R61: And again, I showed these books before, but the book by Ian Deff and John Meet and Alan Erisman, and Jennifer and Mira Slack. This is an open access book that you can just find online, if you want to know the gory details about what's in there, those are good places to start to see what's there. 152 00:22:14.580 --> 00:22:25.289 STFC-RAL-CR03 R61: we think that stuff is great, but it's not just us. So here's a paper from, some people at, Los Alamos, University of Michigan, North West, and Sandia. They… 153 00:22:25.620 --> 00:22:30.490 STFC-RAL-CR03 R61: looked at all the different somber enough smart linear systems that are out there. 154 00:22:30.600 --> 00:22:33.689 STFC-RAL-CR03 R61: The ones in red are the ones that we developed. 155 00:22:33.800 --> 00:22:50.570 STFC-RAL-CR03 R61: And they said, look at a set of easy problems and difficult problems, and looked at what's the best license solver and the best fruit solver for academics. And 3 of these four, we've got a target on parties, to be, there. But it's also worth saying that the licensing terms are very, 156 00:22:50.950 --> 00:22:51.780 STFC-RAL-CR03 R61: Yes. 157 00:22:51.880 --> 00:22:59.639 STFC-RAL-CR03 R61: quite a lot. It costs quite a lot more to, like, improve our solvers. And they've got more effort to develop setting. 158 00:23:01.760 --> 00:23:05.349 STFC-RAL-CR03 R61: So we talked about solving linear algebra, the X equals V. 159 00:23:05.530 --> 00:23:15.350 STFC-RAL-CR03 R61: I've talked about a class called direct methods, so far, but there's also another philosophy for solving new systems, called itcher methods. 160 00:23:15.680 --> 00:23:34.769 STFC-RAL-CR03 R61: So, here's a… here's a cartoon that kind of shows the difference between these two methods. For the direct method, you… you have your matrix, you have your right-hand side, you give it to a computer, and you wait some time. You get no answer back, for… for a period of time, and then all of a sudden, you get the answer, which might be, 161 00:23:35.110 --> 00:23:37.950 STFC-RAL-CR03 R61: Hopefully it gets an older machine precision. 162 00:23:38.480 --> 00:23:41.579 STFC-RAL-CR03 R61: Depending on the condition of the matrix, that might not be the case, but… 163 00:23:41.930 --> 00:23:44.730 STFC-RAL-CR03 R61: But you'll give you the exact answer, let's pretend. 164 00:23:44.990 --> 00:23:52.380 STFC-RAL-CR03 R61: For intuitive methods, there'll be a procedure that gives you successively better and better and better approximations, to the solution. 165 00:23:52.400 --> 00:24:07.729 STFC-RAL-CR03 R61: And if you're solving a problem where you only know the values of your right-hand side for two-digit accuracy, there's no point waiting for 15 digits, because you're getting too much accurate data. So maybe you only want 6 digits of accuracy, and then you can stop it with the… 166 00:24:07.730 --> 00:24:14.720 STFC-RAL-CR03 R61: I should say the sort of this line will change depending on the method that you're using, so sometimes direct will give you 167 00:24:14.860 --> 00:24:16.360 STFC-RAL-CR03 R61: Cool. Accuracy. 168 00:24:16.560 --> 00:24:21.009 STFC-RAL-CR03 R61: sooner than the future. But yeah, that's all network-dependent. 169 00:24:22.840 --> 00:24:27.550 STFC-RAL-CR03 R61: So, here's a, again, the same example that I showed before. It's a 2D, 170 00:24:27.690 --> 00:24:32.989 STFC-RAL-CR03 R61: Buffs on seclusion, disclosing foundments, some of them within the assistant there. 171 00:24:33.100 --> 00:24:41.939 STFC-RAL-CR03 R61: This is using a method called common ingredients, the acquired substance method. You'll get the solution in 2,500 172 00:24:42.120 --> 00:24:44.359 STFC-RAL-CR03 R61: 3.7 or so iterations. 173 00:24:45.830 --> 00:25:03.900 STFC-RAL-CR03 R61: The key to making these things go fast is a technique called preconditioning, which is an accelerator that changes the problem. Simple left preconditioning here. Instead of solving AX equals B, you solve P inverse AX equals P inverse B, so it's a different problem, with different properties, but… 174 00:25:04.130 --> 00:25:11.229 STFC-RAL-CR03 R61: simply happens to converge there. But instead of Solved in, 2,700 iterations. 175 00:25:11.460 --> 00:25:22.129 STFC-RAL-CR03 R61: your conversion rate to iterations. You've got to do some work in here, so each one of these iterations costs more than one of the iterations of the conversion problem. 176 00:25:22.440 --> 00:25:24.780 STFC-RAL-CR03 R61: But you hope. 177 00:25:25.220 --> 00:25:26.430 STFC-RAL-CR03 R61: In the evening. 178 00:25:26.960 --> 00:25:32.799 STFC-RAL-CR03 R61: The aim is the design and recognition number. This is cheap to apply, where the wall for claims is faster. 179 00:25:33.140 --> 00:25:34.070 STFC-RAL-CR03 R61: Dude, yes. 180 00:25:35.600 --> 00:25:49.560 STFC-RAL-CR03 R61: Again, in HSL, we have software, for doing this. There are both the iterative methods, so there's, SynDK, GM, Res, CGLS, IGDSAP, these are all, if you solve the new system, these are all familiar terms, maybe. 181 00:25:49.670 --> 00:25:54.319 STFC-RAL-CR03 R61: But there's also preconditions in here. You can increase trust, precondition of the invoice. 182 00:25:54.430 --> 00:26:01.339 STFC-RAL-CR03 R61: kind of a direct method on drops and pins from, or Algebraic Montego precisioner. 183 00:26:01.680 --> 00:26:03.310 STFC-RAL-CR03 R61: That will, 184 00:26:05.120 --> 00:26:13.760 STFC-RAL-CR03 R61: solve coarser versions of the problem to do something. So, there are both preconditions, acceleration techniques, and the underlying 185 00:26:14.370 --> 00:26:16.369 STFC-RAL-CR03 R61: It should result in the library. 186 00:26:17.700 --> 00:26:30.650 STFC-RAL-CR03 R61: Again, here's the 2D Poisson, equation block we saw earlier. If I add in here, this is just using project gradients, which, as a result, wasn't too great. 187 00:26:30.790 --> 00:26:46.490 STFC-RAL-CR03 R61: But we can… if we do a preconditioner system using algebraic multiple preconditioner on here, then you get similar to the performance of the direct method. It's slightly worse than the soft direct method, takes slightly more time, but roughly equal. 188 00:26:47.530 --> 00:26:52.739 STFC-RAL-CR03 R61: However, this is for a 2D problem. One of the things that kills you with, 189 00:26:53.020 --> 00:27:00.790 STFC-RAL-CR03 R61: one of the things that gets in trouble with, sparse direct methods is it's filling. I said that we ignore zeros. 190 00:27:00.790 --> 00:27:16.779 STFC-RAL-CR03 R61: As you're doing your factorization, some of the zeros will be added in, and it just happens that three-dimensional partial differential equations have the right structure to make that really kind of bite you. So if you do the direct method for three-dimensional partial differential equations, you can find them 191 00:27:16.780 --> 00:27:21.499 STFC-RAL-CR03 R61: quite rapidly, but literally the methods you avoid that, and you can get a nice. 192 00:27:21.610 --> 00:27:31.009 STFC-RAL-CR03 R61: an optimal iterative method, you can do it in older and solve the system in order. Which is really tricky in, real projects. 193 00:27:31.430 --> 00:27:44.830 STFC-RAL-CR03 R61: So, here's an example of one of those real problems. This is work we did with the UK, or are currently doing with the UK Atomic Energy Authority and Los Angeles National Land, and they're modeling, the plasma inside a top map. 194 00:27:45.080 --> 00:28:01.770 STFC-RAL-CR03 R61: And in the problems they want to solve, they're highly anisotropic problems, so if you can see, not much happens in the middle of the top Mac, and you have this one layer in the middle where all the physics happens all of a sudden, so to be able to resolve that, the… 195 00:28:01.770 --> 00:28:18.659 STFC-RAL-CR03 R61: that means the parameters can be up to, 10 to the 10 difference in them, so the… here's, here's the linear system, the PDE that they're solving. The linear system that has to be solved has, an isopropy ratio 196 00:28:19.180 --> 00:28:24.800 STFC-RAL-CR03 R61: We've been testing, so for… the, brown… 197 00:28:25.280 --> 00:28:36.270 STFC-RAL-CR03 R61: one here is a smooth arbitration multiplic preconditioner, and the third one is algebraic multipliers, so these are things that, if you, if it broke down any of this, Pepsi is the, the, the… 198 00:28:36.470 --> 00:28:41.040 STFC-RAL-CR03 R61: packages that everyone uses. These are the state-of-the-art ones in Petsy. 199 00:28:41.280 --> 00:28:49.120 STFC-RAL-CR03 R61: really bad as the anacet property ratio drops, so if this is 10 to the minus 7, it takes, 200 00:28:49.290 --> 00:29:00.669 STFC-RAL-CR03 R61: 40 seconds to reduce the residual by, by an order of magnitude. So we've developed every mission is that, is fairly stable as you go down. 201 00:29:01.050 --> 00:29:06.640 STFC-RAL-CR03 R61: Of course, there's no free lunch, so the complexity of the preconditioner gets worse and worse. 202 00:29:06.710 --> 00:29:24.860 STFC-RAL-CR03 R61: As, as the kind of property increases, property ratio increases, but you couldn't solve these problems before, because it's intractable, so it's more extensive, but, you're going to be solved. Are you using an adaptive grid method here? 203 00:29:25.390 --> 00:29:30.729 STFC-RAL-CR03 R61: I think there is, yes, yes, underneath it all. So, yeah, we've left that to the, 204 00:29:31.040 --> 00:29:39.690 STFC-RAL-CR03 R61: the physicists come up with a great thing with us, the linear systems, and, yeah, and we've developed methods to solve that. But yeah, it's a… 205 00:29:40.420 --> 00:29:42.720 STFC-RAL-CR03 R61: The grid is, is, 206 00:29:43.370 --> 00:29:48.920 STFC-RAL-CR03 R61: find out where it needs to be fined at, and they've got some feedback mechanism in there to look at where it goes. 207 00:29:51.860 --> 00:30:07.510 STFC-RAL-CR03 R61: there's tons of other stuff that goes on in the branch. I'm not going to talk about these, but just to say some, yeah, these are some… if any of these trigger any, things that you… please come talk to us. CUR decomposition, agility problems. 208 00:30:08.460 --> 00:30:09.529 STFC-RAL-CR03 R61: Holy sense. 209 00:30:10.670 --> 00:30:19.600 STFC-RAL-CR03 R61: Okay, now I'm gonna move on to the continuous optimization. So in optimization, There's minimizing… 210 00:30:19.800 --> 00:30:36.449 STFC-RAL-CR03 R61: function of X. So given a function, possibly subject to some constraints, what's the value that minimizes this function? So can you find this point here, or even this point here, to be a local minimum? This is the rule. So how do you do that? 211 00:30:36.880 --> 00:30:39.030 STFC-RAL-CR03 R61: identity with them. 212 00:30:39.290 --> 00:30:41.699 STFC-RAL-CR03 R61: WebEx is meetings. 213 00:30:42.580 --> 00:30:45.390 STFC-RAL-CR03 R61: So we have software, called Gallaghett. 214 00:30:45.870 --> 00:30:58.359 STFC-RAL-CR03 R61: And Canada has a library of full-time packages, with C, Pack, and Julia Mathline interfaces, so the portrait, or one of your favorite languages, hopefully, 215 00:30:59.350 --> 00:31:00.440 STFC-RAL-CR03 R61: supported? 216 00:31:02.210 --> 00:31:06.670 STFC-RAL-CR03 R61: We also have other software, so, 217 00:31:06.810 --> 00:31:17.280 STFC-RAL-CR03 R61: there's sources of problems, and some other things, and, and here's optional is forgetting that, you know, part of all the other software and, and mix-written, 218 00:31:17.510 --> 00:31:27.669 STFC-RAL-CR03 R61: The book on press region methods, if that means anything, and, and complexity of algorithms for non-convex optimization, so you'll go and look at these. 219 00:31:27.800 --> 00:31:28.650 STFC-RAL-CR03 R61: So… 220 00:31:28.890 --> 00:31:42.370 STFC-RAL-CR03 R61: Optimization in general, in general, we're interested. I'm going to focus on one specific problem here, and that's not a least squares problem. So, how do we minimize a sum of squares subject to some constraints? 221 00:31:42.790 --> 00:31:45.910 STFC-RAL-CR03 R61: So… 222 00:31:46.370 --> 00:32:04.329 STFC-RAL-CR03 R61: A few years ago, probably 8 or 2 years ago, I think time goes by quickly, guys in ISIS, the CBU team, when they came forward, is the mounted, developers were working with us from the CBU. They had this data set that came, from the instrument. 223 00:32:05.190 --> 00:32:16.679 STFC-RAL-CR03 R61: In Manta, they were using the Blue Scientific Library, and they asked, can you fit… they were trying to fit two Gaussians with some background noise, some, a background function. 224 00:32:17.070 --> 00:32:25.610 STFC-RAL-CR03 R61: linear function to this, and GSLs give them, this solution, which doesn't fit the data at all. 225 00:32:25.840 --> 00:32:32.809 STFC-RAL-CR03 R61: And they came to us and said, can you tell us why this doesn't fit? And, can you give us a method that does fit, and they fit? 226 00:32:33.130 --> 00:32:36.160 STFC-RAL-CR03 R61: So, we developed, this software route fit. 227 00:32:36.670 --> 00:32:43.490 STFC-RAL-CR03 R61: Rapid, gives you the red line that you did expect to go through here. Essentially. 228 00:32:44.250 --> 00:32:58.649 STFC-RAL-CR03 R61: GSL was using the Markov method, which uses first-order derivatives, but nothing higher than that. We developed Market that, if you have high-order derivatives, you can work out the Hessian above, and then it can use that information. 229 00:32:58.940 --> 00:33:00.769 STFC-RAL-CR03 R61: Give you a benefit because of that. 230 00:33:00.940 --> 00:33:03.740 STFC-RAL-CR03 R61: So we give them outfit. 231 00:33:04.240 --> 00:33:07.750 STFC-RAL-CR03 R61: They're very happy with that, and that's the final advantage. 232 00:33:08.060 --> 00:33:14.979 STFC-RAL-CR03 R61: Again, this is a label drip that we built for ISIS, but we made it available for all, and so… 233 00:33:15.230 --> 00:33:30.460 STFC-RAL-CR03 R61: use the Romford package, it doesn't just do some squares, you can do weights, and we can do some origination on this one. It was picked up by the NAG Library, Nagler company, based in Oxford, so if, if you look at their, 234 00:33:31.090 --> 00:33:43.310 STFC-RAL-CR03 R61: solving monumently squares, this is what's underneath there, solver there. And, more recently, AMT, the chip manufacturer, has taken this method, and that's what they use for their data analytics library. 235 00:33:43.820 --> 00:33:57.459 STFC-RAL-CR03 R61: Trying to show the publicity that NAG did. So they asked their data, but it's not solder, this is their, blog post announcing this, and they say, the new SOLBA, EO4GG, is what we call Barbit. 236 00:33:58.250 --> 00:34:01.670 STFC-RAL-CR03 R61: All four different ways of calling things that don't meet a certain number of needs. 237 00:34:01.850 --> 00:34:07.959 STFC-RAL-CR03 R61: The new song offers unprecedented robustness and a significant speed up all the current alternatives in the library. 238 00:34:08.130 --> 00:34:23.010 STFC-RAL-CR03 R61: Namely E04 GP, the old one for fan-constrained boundary, particularly square solvers, and E04US for problems with simple variable problems. They're highly recommended to upgrade to the new solver, so this is a solver we 239 00:34:23.389 --> 00:34:30.829 STFC-RAL-CR03 R61: There's a crisis, and when it's used, you know, it's used all around the world. We use Maggie and G-Soft software. 240 00:34:32.489 --> 00:34:38.590 STFC-RAL-CR03 R61: Ralph, it gives you a local minimum, so it will give you one minimum, 241 00:34:39.540 --> 00:34:46.469 STFC-RAL-CR03 R61: There are also problems… so another, lead magnetizes had the problem that they wanted a global movement, so… 242 00:34:47.770 --> 00:34:49.730 STFC-RAL-CR03 R61: If you're doing Crystal Feel Fitted. 243 00:34:50.770 --> 00:35:06.099 STFC-RAL-CR03 R61: If you go to the BMI scientist, the lead scientist of the instrument, and you give them the data, they'll say, you start at this starting point, and you'll get a solution, and you run this through file fit, or GSL, whatever, that'll give you the fit that you'd expect. 244 00:35:07.680 --> 00:35:14.859 STFC-RAL-CR03 R61: If you perturb the value that the Beeline scientists give you by, I think it was 10 to the minus 3, so not very much. 245 00:35:15.240 --> 00:35:31.709 STFC-RAL-CR03 R61: you will get another fit, which is a local minimum, so it's a valid, this is a valid solution. It is a local minimum, but it isn't the minimum that the physicists will be looking for, and that it does… it misses these peaks. 246 00:35:31.850 --> 00:35:38.659 STFC-RAL-CR03 R61: So, they said, can you give us a global? And I'm using the global, you know, yeah. 247 00:35:39.230 --> 00:35:43.179 STFC-RAL-CR03 R61: Again, this, this was funded through, the Lunday center. 248 00:35:43.580 --> 00:35:52.799 STFC-RAL-CR03 R61: So, Gary, Mick, and Megan did some work on this, doing a, global multi-start method, so you essentially 249 00:35:52.930 --> 00:35:56.429 STFC-RAL-CR03 R61: Get a grid, start in lots of different points, and, and then… 250 00:35:56.980 --> 00:35:59.639 STFC-RAL-CR03 R61: Refine the grid as you go down to make sure that 251 00:35:59.820 --> 00:36:17.200 STFC-RAL-CR03 R61: you might not get the broader minimum, but you've got a better chance of getting broader minimum. And, yeah, so the lighter is better, and, for the five example problems they gave us, gives us what's in for the multi-staff, this was all the methods that were implemented and mounted at that time. 252 00:36:17.230 --> 00:36:23.919 STFC-RAL-CR03 R61: And, yeah, this method is like. It does what we need. And, 253 00:36:24.320 --> 00:36:38.959 STFC-RAL-CR03 R61: Again, it's a LEGO brick. It was built for, crystal things fit in in the mountain, but it's a package that you can sold, golf it, and you can run it on any problem. 254 00:36:39.450 --> 00:36:57.259 STFC-RAL-CR03 R61: Yeah, so ISIS were delighted with this. The problem they had was a postdoc wouldn't have to do some data analysis. You needed to wait and have time with the VMind scientists to tell you where to start, otherwise you just lose that data. You won't get an answer out. And so this is more expensive to run, but you're, 255 00:36:57.820 --> 00:36:58.800 STFC-RAL-CR03 R61: in my… 256 00:36:58.950 --> 00:37:06.670 STFC-RAL-CR03 R61: take a couple of hours to run, whereas this might take a couple of minutes, but the couple hours running is my fault that it's sitting on the system, not… 257 00:37:07.310 --> 00:37:08.960 STFC-RAL-CR03 R61: I'm not being used at all. 258 00:37:11.180 --> 00:37:20.900 STFC-RAL-CR03 R61: So these are a few fitting problems, that we've done, and when we developed GoFit and RouteFit, again, the guys in ISIS were very sensibly, we've been mute 259 00:37:21.080 --> 00:37:34.510 STFC-RAL-CR03 R61: five problems, or one problem, and this does well in there. How do we know that this is better over all problems? Out of that question came, another, software package that we, well, the benchmark in. 260 00:37:34.510 --> 00:37:42.949 STFC-RAL-CR03 R61: the idea behind fit benchmarking is that for any source of test problems, and so these are ones that we currently have in fit benchmarking, there's, 261 00:37:42.980 --> 00:37:58.849 STFC-RAL-CR03 R61: the SASFIT that's, built by, the Culture Institute, Switzerland, mounted at ISIS, SASFIT from Simon, the NIST standard problem, so for anything, you take the actual data, the actual models from these problems. 262 00:37:59.070 --> 00:38:11.930 STFC-RAL-CR03 R61: hook them into all the different versions of, data fitting software. So, I guess Mini comes from the article visits community, so probably people in this room know that well. 263 00:38:12.750 --> 00:38:20.319 STFC-RAL-CR03 R61: pieces from Facebook, there's our goal fit and our outfit, there's Ceres from Google, AMD stuff, we've got a… 264 00:38:20.320 --> 00:38:32.989 STFC-RAL-CR03 R61: whatever's out there. What tends to happen is that the community will use the software that comes from that community, and not be aware of all the other things that are out there. This is a way that people can easily try things from other communities. 265 00:38:33.470 --> 00:38:41.960 STFC-RAL-CR03 R61: If it doesn't work, it doesn't work, and you just move on, but you can highlight, this is a promising method, let's spend the time properly and integrate it into my software. 266 00:38:42.080 --> 00:38:44.669 STFC-RAL-CR03 R61: And, and get back to work. 267 00:38:45.160 --> 00:38:59.159 STFC-RAL-CR03 R61: So we've seen one of these tables before. If anyone gives you these tables, so if you fit in a load of problems. Here are a load of different, solutions, so it's like ILS, RAFID, Legmark, GSL in this case. 268 00:38:59.260 --> 00:39:11.679 STFC-RAL-CR03 R61: The top color's the relative accuracy, bottom colors are relative runtime. It's really easy to have a method that runs really quickly and gives you the completely wrong answer, so you need to have both of these to know if it's a good method or not. 269 00:39:12.000 --> 00:39:13.629 STFC-RAL-CR03 R61: You see here there's lots of… 270 00:39:14.140 --> 00:39:25.189 STFC-RAL-CR03 R61: Orange light is better, lots of orange, or lots of light colors around this, blue light mark method, but it's purple at the top, which means the accuracy isn't bright. Oops. 271 00:39:27.210 --> 00:39:37.689 STFC-RAL-CR03 R61: And, yeah, you can print on there and give you all the information you want, and you see that the fit is just completely wrong. And you want to avoid these methods on your dataset. These are all 272 00:39:37.830 --> 00:39:51.389 STFC-RAL-CR03 R61: These are all there because they work with some sort of data, and finding which method it works well on your data can be hard. And you can look all the way as well and see, for this problem, these are the passive methods that work. 273 00:39:52.690 --> 00:39:57.720 STFC-RAL-CR03 R61: You have… Performance profiles, which is a good way of checking. 274 00:39:59.010 --> 00:40:03.950 STFC-RAL-CR03 R61: All the methods in the top right corner, top left corner are the best methods. 275 00:40:04.700 --> 00:40:19.359 STFC-RAL-CR03 R61: And this is one place that we'll be working with the computational community and theoretical and experimental partners. I put the names of the people up in this… I don't know if anyone in this room knows any of the people in here, but if you do, 276 00:40:19.680 --> 00:40:21.389 STFC-RAL-CR03 R61: These are the people we're working with. 277 00:40:21.630 --> 00:40:26.599 STFC-RAL-CR03 R61: One of the things… it's the theoretical part of this, community. 278 00:40:27.540 --> 00:40:32.680 STFC-RAL-CR03 R61: Their bottleneck is fast linear systems, and they have a need for solving, 279 00:40:33.020 --> 00:40:38.520 STFC-RAL-CR03 R61: Not maybe space problems, and so we'll be hooking some of their software into the benchmarking framework. 280 00:40:38.730 --> 00:40:51.250 STFC-RAL-CR03 R61: Interestingly, they don't seem to use… I don't know the split between experimental and theoretical, but they seem unworth that there's a community that already does this. The people that develop their, their, 281 00:40:51.720 --> 00:40:58.969 STFC-RAL-CR03 R61: at least Squaresolvers are retiring. It's a problem that software isn't maintained forever if it's running in the universe. 282 00:40:59.130 --> 00:41:03.259 STFC-RAL-CR03 R61: So we're seeing what's… what's the next thing that we can add in. 283 00:41:05.600 --> 00:41:11.999 STFC-RAL-CR03 R61: And last but not least is the inverse problems group. And again, Margaret is there, and… 284 00:41:12.880 --> 00:41:21.129 STFC-RAL-CR03 R61: many of these slides will be familiar to Marcus, if you have any questions, you can ask Mark. So, start with what is an inverse problem? So… 285 00:41:22.090 --> 00:41:25.530 STFC-RAL-CR03 R61: Thank you, yeah. Congrat? 286 00:41:25.820 --> 00:41:29.480 STFC-RAL-CR03 R61: You take a lot of x-rays of the kindred. 287 00:41:29.590 --> 00:41:34.719 STFC-RAL-CR03 R61: And fit the results on the detector, and he did it round. And the question is, what's inside? 288 00:41:35.500 --> 00:41:41.310 STFC-RAL-CR03 R61: The egg, and you can kind of see from the pictures, the pictures of the family in there. 289 00:41:42.270 --> 00:41:46.060 STFC-RAL-CR03 R61: And you can solve, the difference. So the noisy data would be 290 00:41:46.280 --> 00:42:01.330 STFC-RAL-CR03 R61: the pictures that come at the end. The unknown is what's the model that's in front of it, and then you assume you have the model, the forward model, which might be linear or might be non-linear, of what's the reverse. So this is the problem. 291 00:42:02.680 --> 00:42:04.780 STFC-RAL-CR03 R61: And, in… 292 00:42:05.170 --> 00:42:14.519 STFC-RAL-CR03 R61: You do the… solve this mathematical problem, you get a reconstruction, you get some slices through the data, and then you can do the next step and pick out 293 00:42:14.570 --> 00:42:25.370 STFC-RAL-CR03 R61: Do the segmentation in here and say, this is… there's a skateboard, there's a little cartoon death character, here's the thing, you know, the chocolate, here's… and then… 294 00:42:25.430 --> 00:42:27.089 STFC-RAL-CR03 R61: See all the bits inside it. 295 00:42:27.760 --> 00:42:41.539 STFC-RAL-CR03 R61: recently had to visit the dentist. It was amazing. They did this CT on my, you put your head in, I think, they go around, and then within a couple of minutes, 10 minutes, they have this reconstruction of your teeth. 296 00:42:41.540 --> 00:42:48.310 STFC-RAL-CR03 R61: right there in dental surgeries, it's all people, like, moderate developing algorithms behind that, that makes this all good. 297 00:42:49.080 --> 00:42:51.299 STFC-RAL-CR03 R61: Of course, this is… this is… 298 00:42:51.400 --> 00:43:03.990 STFC-RAL-CR03 R61: much as hard to do, this is impossible to do in many cases, so this is… this guy's, Haramada, a mathematician, who says, this is an ill-sposed problem in case there's either no solution, the solution is unique. 299 00:43:04.020 --> 00:43:15.300 STFC-RAL-CR03 R61: Or if you're solution is sensitive to noise, so you have to explain the formation of your problems and regularization, we're able to make sure that these aren't, a problem for your reconstruction. 300 00:43:17.040 --> 00:43:23.310 STFC-RAL-CR03 R61: So, some examples, and again, his, Margaret's work, 301 00:43:23.810 --> 00:43:30.440 STFC-RAL-CR03 R61: One challenge in here is mortifying compensation, so if you think of a CT scan. 302 00:43:30.650 --> 00:43:41.679 STFC-RAL-CR03 R61: If your body is in a CT scanner, you're told to stay as still as you can, but you're obviously moving, you breathe it, if you have some heart and lens, it's moving. As you give the scan round. 303 00:43:41.920 --> 00:43:52.689 STFC-RAL-CR03 R61: If you try to do the reconstruction, you'll have compensation for the motion, and you'll get a blurry image, as you can imagine. There are techniques you can use to do the motion compensation. 304 00:43:56.000 --> 00:43:58.020 STFC-RAL-CR03 R61: a nice sharp image. 305 00:43:58.400 --> 00:44:10.749 STFC-RAL-CR03 R61: What Margaret and Leticia, the collaborators, did, you, you take, so, FISTA and the primordial hybrid algorithm, algorithms that are out there in the wild. 306 00:44:10.940 --> 00:44:17.589 STFC-RAL-CR03 R61: These, if you run them for long enough, sent to Epo, so you run through the data centers, you'll get a nice sharp image. 307 00:44:19.740 --> 00:44:28.189 STFC-RAL-CR03 R61: what they've done is kind of the stochastic frame will do… so if you add some stochasticity into it, you can get a nice, sharp image at the 10 eBucks, so… 308 00:44:28.880 --> 00:44:40.249 STFC-RAL-CR03 R61: Rather than waiting for 70 minutes in the doctor's office for getting your results, you can wait for 10 minutes in the doctor's office to get your results, and that makes a big difference. 309 00:44:40.880 --> 00:44:47.270 STFC-RAL-CR03 R61: And, yeah, so here's the paper where the details are in, so the teacher and Albert who responds to. 310 00:44:47.440 --> 00:44:49.030 STFC-RAL-CR03 R61: And it's a QR code. 311 00:44:49.200 --> 00:44:50.630 STFC-RAL-CR03 R61: deceiving. 312 00:44:51.860 --> 00:44:54.479 STFC-RAL-CR03 R61: Another envious problem that we've been, 313 00:44:55.100 --> 00:45:08.510 STFC-RAL-CR03 R61: working a lot with is, typography. So this is a technique that's used in… in Diamond, the central laser facility, and there was the Environment Institute, so the people that do this then. So here, you have… say you have an x-ray. 314 00:45:08.950 --> 00:45:19.619 STFC-RAL-CR03 R61: And you fire an X-ray at a sample, move the sample a little bit, fire again, move a bit with overlapping images, so you have a diffraction pattern at the back. 315 00:45:21.170 --> 00:45:34.350 STFC-RAL-CR03 R61: you assume that your sample is thin, then you just tick the Fourier transform of the exit wave, and you try… it's, again, a non-release squares problem. A very big, very complicated non-lease squares problem to try and match the data there. 316 00:45:35.430 --> 00:45:41.139 STFC-RAL-CR03 R61: The standard method in the Fuji is this ePay algorithm, 317 00:45:41.860 --> 00:45:49.430 STFC-RAL-CR03 R61: We've looked at one of the methods, one of the optimization methods out there, that you can use them to review this, hopefully. 318 00:45:49.740 --> 00:45:55.440 STFC-RAL-CR03 R61: low-rank, PHGS method. 319 00:45:56.810 --> 00:45:58.549 STFC-RAL-CR03 R61: And these GIF, yeah. 320 00:45:59.050 --> 00:46:04.779 STFC-RAL-CR03 R61: Better reconstructions in some cases, or at least another tool in your toolbox that you can use. 321 00:46:07.920 --> 00:46:12.320 STFC-RAL-CR03 R61: Another question, so that previous stuff would assume that the sample is 322 00:46:13.030 --> 00:46:30.710 STFC-RAL-CR03 R61: thing. In the real world, things have volume, things are 3D, 3-dimensional. What can you do in that case? Standard technique is multi-slice, so you assume that there's multiple slices through your sample, which would work very well if you have crystalline structures, if you do it through salt or something. 323 00:46:32.380 --> 00:46:39.050 STFC-RAL-CR03 R61: something that has a nice crystallized structure inside, this is a good approximation. If you have some biological sample, 324 00:46:39.270 --> 00:46:44.329 STFC-RAL-CR03 R61: Yeah, it doesn't work so well. It's not as well approximated by the slices. 325 00:46:44.480 --> 00:46:47.950 STFC-RAL-CR03 R61: And you also have some… 326 00:46:48.130 --> 00:46:52.519 STFC-RAL-CR03 R61: You can't parallelize something that's inherently, so it's going to take the slice to get slice. 327 00:46:53.260 --> 00:46:56.290 STFC-RAL-CR03 R61: So how do you… how do you agree with the examples? 328 00:46:57.580 --> 00:47:02.610 STFC-RAL-CR03 R61: And so what we've been helping with here is, well, actually, you can solve 329 00:47:03.470 --> 00:47:17.779 STFC-RAL-CR03 R61: the X-ray going in is just a wave, so if you can solve for wave equation within the sample, then take that as a wave and give the reconstruction back, you can get better reconstructions. So the right approximation that you use here is the fract wave equation. 330 00:47:18.380 --> 00:47:31.580 STFC-RAL-CR03 R61: And that brings us nicely back to where we started, that this is a really difficult linear system to solve, and, you need to have good preconditions, good, intricate methods to solve this well, and so that leads to the… 331 00:47:31.830 --> 00:47:40.040 STFC-RAL-CR03 R61: better the outreach methods that can enable the stuff in Diamond, than to do the thing we need to do. 332 00:47:42.540 --> 00:47:49.049 STFC-RAL-CR03 R61: That's where I'm gonna leave it. Yeah, so I'll say that. I'm the… Then we did computational maths. 333 00:47:49.750 --> 00:47:51.070 STFC-RAL-CR03 R61: We're interested in 334 00:47:51.180 --> 00:48:01.490 STFC-RAL-CR03 R61: Linear algebra, optimization, inverse problems, but we also have interest in differential equations, numerical integration, time-stepping methods, all sorts of things in common math, so… 335 00:48:01.630 --> 00:48:05.210 STFC-RAL-CR03 R61: Yeah. Can't focus with anything you can add. 336 00:48:05.580 --> 00:48:09.800 STFC-RAL-CR03 R61: We're always on the lookout for some problems, and it's a… 337 00:48:11.260 --> 00:48:16.049 STFC-RAL-CR03 R61: you guys do interesting stuff. If there's something that you wish they would see. 338 00:48:16.050 --> 00:48:31.390 STFC-RAL-CR03 R61: If only we had a method that could do this faster, or could do this more accurately, those are the kind of things that are really interesting to us, and if it's not of interest to us, then there's someone in the community that we belong to, that I'm sure will be interested as well, and I'd be very happy to connect you up with people. 339 00:48:31.720 --> 00:48:43.210 STFC-RAL-CR03 R61: Again, memorabo.amerald.ncwk is the website. There's my email address, and we're on the bottom corridor of Bar 71, so as you walk off for lunch. 340 00:48:43.360 --> 00:48:46.189 STFC-RAL-CR03 R61: Yeah, always pop in and say hi if you want to see. 341 00:48:53.200 --> 00:48:55.940 STFC-RAL-CR03 R61: Thank you very much. Have a nice talk. 342 00:48:57.020 --> 00:49:02.890 STFC-RAL-CR03 R61: Yes, please. Thank you very much. I'm wondering how the… 343 00:49:03.300 --> 00:49:07.540 STFC-RAL-CR03 R61: licensing policy evolved. For example, for SSL, 344 00:49:08.060 --> 00:49:20.590 STFC-RAL-CR03 R61: how was it in the beginning, and now that you have a very formulated, customized license, it's not particularly standard open license on it. So how does that work? Who do you have to work with for this? 345 00:49:21.270 --> 00:49:24.100 STFC-RAL-CR03 R61: So on, yeah. Yes, well, 346 00:49:25.650 --> 00:49:42.169 STFC-RAL-CR03 R61: it's a complicated… so it's bid… the people in bid, the innovations team in bid are the people who… within SDFC that we work with. The reason why the licensing is as it is, is for complicated historical reasons, but, as I said, the library started at the Harbor Labs. 347 00:49:42.520 --> 00:49:55.010 STFC-RAL-CR03 R61: the IP there was privatized, so it went out to a company called, I think it was AEA Technologies Limited, but it was then bought by another company, it was bought by another company. So, everything pre-1990, the IP is owned by 348 00:49:55.210 --> 00:50:02.919 STFC-RAL-CR03 R61: big company. And I think post-1990, we all might be off, you know, known as the idea of the Who while we're all together. 349 00:50:03.030 --> 00:50:04.240 STFC-RAL-CR03 R61: So… 350 00:50:04.500 --> 00:50:11.130 STFC-RAL-CR03 R61: We have… we've got an agreement, that has a standard set of licenses, so we had to… 351 00:50:11.500 --> 00:50:23.850 STFC-RAL-CR03 R61: we argued very hard to make the academic licenses free, which wasn't… well, first of all, it was, UK academics could have it for free, then we managed to expand that out to be for all academics could have it for free. 352 00:50:24.120 --> 00:50:34.089 STFC-RAL-CR03 R61: But yeah, it's the shackles of this license agreement. I mean, negotiations are big now. Well, obviously, everything… everything past 1990, but it's now 36 years ago. 353 00:50:34.090 --> 00:50:47.519 STFC-RAL-CR03 R61: We own the copyright of. So we have a version of the library which is completely copyrighted CFC, which is everything that anyone actually uses these days, unless you're trying to resurrect some full-time God from 1970, or whatever. 354 00:50:47.700 --> 00:50:51.380 STFC-RAL-CR03 R61: Yes, I'm in the final stages of trying to 355 00:50:51.710 --> 00:50:55.500 STFC-RAL-CR03 R61: Get some license templates out. 356 00:50:56.630 --> 00:50:58.739 STFC-RAL-CR03 R61: There is… we do have to have 357 00:50:59.150 --> 00:51:18.069 STFC-RAL-CR03 R61: The only… the only thing that pays for it is ISIS and eCamp at the moment, so if we… if it was completely… well, it could be completely open source, but we need some sort of way to then get donations in, or to sell consultancy to… to be able to pay for future development, the future self. There's complications in… So, I mean, in the very beginning. 358 00:51:18.070 --> 00:51:21.810 STFC-RAL-CR03 R61: it must have been sort of an open license. Later on, it was… 359 00:51:22.070 --> 00:51:40.059 STFC-RAL-CR03 R61: sort of closed, or… I mean, do you remember? Because in academics, we tend to keep all our, you know, broadcasts or whatever that we create as a, you know, open source and so on. So, in that, how that change happened, that's something which… 360 00:51:40.810 --> 00:51:49.729 STFC-RAL-CR03 R61: I'm wondering… Well, I think software licensing meant a very different thing in the 1960s than it does today, and that's the… it was very much… 361 00:51:49.960 --> 00:51:55.299 STFC-RAL-CR03 R61: Here's a CD that we ordered from this evening, that we will deliver, or… 362 00:51:55.480 --> 00:51:58.159 STFC-RAL-CR03 R61: I'll just say a bunch of other side of the day. 363 00:51:58.280 --> 00:51:59.070 STFC-RAL-CR03 R61: And… 364 00:51:59.450 --> 00:52:12.500 STFC-RAL-CR03 R61: Yeah, so I don't know what the original… the terms of the original license was. I've only seen the ones post-1990, and these are bespoke agreements that are… yeah, are not in the standard open source. 365 00:52:13.510 --> 00:52:17.589 STFC-RAL-CR03 R61: Thank you, it was a really pleasant presentation, thanks. 366 00:52:18.940 --> 00:52:30.950 STFC-RAL-CR03 R61: I was gonna ask, how do you get funded? As in, is it largely, sort of metal left, SDHC money coming down, or do you… 367 00:52:31.160 --> 00:52:36.419 STFC-RAL-CR03 R61: If we wanted to do something with you, for example, would we have to get project funding in, and then… 368 00:52:36.680 --> 00:52:45.369 STFC-RAL-CR03 R61: pay for your time with overheads, or what's… what's the sort of model? Yes, good question. Okay. There's, 369 00:52:45.620 --> 00:52:47.810 STFC-RAL-CR03 R61: Yeah, so there's the problems… 370 00:52:47.980 --> 00:52:57.190 STFC-RAL-CR03 R61: If they make sense, then that's a… that's a way to do it. So then there's the… the COSEC T10, theoretical and experimental part of this. 371 00:52:57.450 --> 00:53:01.520 STFC-RAL-CR03 R61: Is it easy for us to work with the platform because it's community. 372 00:53:01.720 --> 00:53:13.300 STFC-RAL-CR03 R61: to do that. We, yeah, apply for research grants, so there's possible that people 373 00:53:13.540 --> 00:53:16.100 STFC-RAL-CR03 R61: joint research advances at DWE Forward. 374 00:53:18.060 --> 00:53:28.460 STFC-RAL-CR03 R61: Yeah, it depends how close our current things are. Like, each of our brings internal income, so if there's a… if there's a need to do, 375 00:53:28.820 --> 00:53:43.910 STFC-RAL-CR03 R61: better than your solvers, applies to class statistics, but we would probably just work on that with you, given something that's completely optional. Yeah, I think chatting will work out what's the best route to get things done. For me, there's no kind of 376 00:53:44.180 --> 00:53:49.520 STFC-RAL-CR03 R61: core funding to do what we like in Italy. That's… that's obviously the hard part. 377 00:53:52.870 --> 00:53:57.710 STFC-RAL-CR03 R61: Okay, sorry, it was, so, I think… Sweet. 378 00:53:57.980 --> 00:53:59.580 STFC-RAL-CR03 R61: They have this experiment. 379 00:53:59.740 --> 00:54:01.190 STFC-RAL-CR03 R61: That's San Jose. 380 00:54:01.350 --> 00:54:04.780 STFC-RAL-CR03 R61: HSL an apology, because we did have… 381 00:54:05.310 --> 00:54:11.660 STFC-RAL-CR03 R61: some Matrix inversion in there until 2017, on there. And… 382 00:54:12.620 --> 00:54:27.339 STFC-RAL-CR03 R61: Then people freaked out about it. But we replaced it with something open source, so… sorry about that. Although, I suspect it could have been covered by an academic license, but then we had a problem of re-licensing it. We only sorted out our actual licensing for our software around that. 383 00:54:27.600 --> 00:54:29.130 STFC-RAL-CR03 R61: So we have a, 384 00:54:29.940 --> 00:54:34.639 STFC-RAL-CR03 R61: Yeah, we have a version of it that's closed source that, still has that history as well, most of it. 385 00:54:38.330 --> 00:54:51.889 STFC-RAL-CR03 R61: The… yeah, the other thing I was going to ask is, do you… so you've got quite a close collaboration with AMD. Do you run… do you have any stuff that… do you run any stuff on AMD GPUs? And I guess more general question about what sort of hardware you have access to and do… 386 00:54:52.760 --> 00:54:58.530 STFC-RAL-CR03 R61: So we, the AMD GPU stuff, we are… in… 387 00:54:59.710 --> 00:55:02.829 STFC-RAL-CR03 R61: They are interested in working with us, the developer, man. 388 00:55:03.370 --> 00:55:11.989 STFC-RAL-CR03 R61: Okay, they have, CTS, that's the solver. They want to add the punches there, so that we are… 389 00:55:12.120 --> 00:55:13.710 STFC-RAL-CR03 R61: Yeah, inactive. 390 00:55:13.940 --> 00:55:15.740 STFC-RAL-CR03 R61: Talks with them, to… to… 391 00:55:16.390 --> 00:55:27.350 STFC-RAL-CR03 R61: HSR technology and the aspect of it, but licensing and stuff is obviously the tricky point there on both sides. But yeah, we were talking with them. 392 00:55:27.580 --> 00:55:33.289 STFC-RAL-CR03 R61: browser, that's NVIDIA, GPUs that targets. 393 00:55:34.450 --> 00:55:35.809 STFC-RAL-CR03 R61: Yeah, I should have read some. 394 00:55:36.180 --> 00:55:39.319 STFC-RAL-CR03 R61: Yeah, they use CPU plus GPUs that it can get. 395 00:55:39.840 --> 00:55:46.660 STFC-RAL-CR03 R61: It's like 6 times speed up is what we've seen for that, if you needed to do that. 396 00:55:47.630 --> 00:55:51.669 STFC-RAL-CR03 R61: That's fully open source, so you can use that in Zoom or that in. 397 00:55:52.160 --> 00:56:08.130 STFC-RAL-CR03 R61: Yeah. Do you know what you switched to? You did? Eigen. Okay. Just because we were switching to Eigen everywhere. So, it was just one specific bit of reconstruction code that had that. It was a bit of a mess, and we solved that at the same time. 398 00:56:10.270 --> 00:56:11.340 STFC-RAL-CR03 R61: That's true. 399 00:56:12.350 --> 00:56:20.660 STFC-RAL-CR03 R61: Wasn't… wasn't my decision to put it in. This was dated back to 2004, something, something like that, before I would start… started university. 400 00:56:23.780 --> 00:56:42.770 STFC-RAL-CR03 R61: That's very interesting, actually. I'm amazed at how big it is now, and we don't really know about it, because in the past, I interacted with people like Ian Duff in Sparse Matrix stuff, but what… the question I really want to know now is, a lot of departments are developing their own simulation codes. 401 00:56:43.150 --> 00:56:45.690 STFC-RAL-CR03 R61: And do you get involved with those? 402 00:56:46.260 --> 00:56:53.909 STFC-RAL-CR03 R61: Or do you actually just leave the departments alone? Because, we're developing simulation codes in CLF, 403 00:56:53.990 --> 00:57:12.179 STFC-RAL-CR03 R61: in rail space, and also in technology. And these are kind of codes that… you either get them off the shelf, you know, or you develop them in-house. The ones that you develop in-house, I… the ones that I know about, we don't interact a lot with the combination of critic group. 404 00:57:12.890 --> 00:57:21.060 STFC-RAL-CR03 R61: So how can we increase that? I mean, it goes back to Jen's point of question. How do we increase the interaction? 405 00:57:21.170 --> 00:57:30.829 STFC-RAL-CR03 R61: Yes, because I can give you a few questions, a few problems now, but then, do I have the money to pay for it? No, I don't. 406 00:57:31.680 --> 00:57:37.200 STFC-RAL-CR03 R61: Good question. Yes, we, we… We do. 407 00:57:37.790 --> 00:57:53.040 STFC-RAL-CR03 R61: where we have it tends to be where things go wrong, and then it really is… what I'd really like to see is us being involved in a conversation when these things have been started out, so we have… there's a COVID particle physics that they're putting together. 408 00:57:53.040 --> 00:58:08.659 STFC-RAL-CR03 R61: Which doesn't… they have some numerical instabilities, and they're trying to sort out a way which numerical instabilities have gone, so they've come to us. We weren't in their original plan, so we've got the money to do that. We've gone to the lovely center, CLF is in this collaboration. 409 00:58:08.880 --> 00:58:21.030 STFC-RAL-CR03 R61: they gave us a bit of money to help them and make that work, which was… which is good. Now they're moving from 2D code to the 3D piece, and architecture all from the start again. 410 00:58:21.430 --> 00:58:31.319 STFC-RAL-CR03 R61: Elc are kind of unwilling to fund things, because it's not really have been the researchers that use the facility, it's a researcher court, but there's a research brand that pays for this. 411 00:58:31.680 --> 00:58:33.880 STFC-RAL-CR03 R61: It all gets messy and muddy. 412 00:58:34.130 --> 00:58:35.669 STFC-RAL-CR03 R61: I think it's a good thing to… 413 00:58:36.520 --> 00:58:45.959 STFC-RAL-CR03 R61: When you hear Ian Chapman talk, it's like everyone should all work together. I think it's a good time to… to say, well, give us some money to all work together. There should be some spending to make this happen. 414 00:58:46.170 --> 00:58:48.130 STFC-RAL-CR03 R61: We may be in a different bucket. 415 00:58:48.130 --> 00:59:08.719 STFC-RAL-CR03 R61: That's it. We could get developers from across departments to meet and talk to each other. I mean, in the past, I mean, it goes back to a problem that we're working on, still working on, I mean, we used to interact strongly with Particle Physics Group, particularly with, 416 00:59:08.720 --> 00:59:13.339 STFC-RAL-CR03 R61: Say, neutrino physics, and associated with that, supernova explosions. 417 00:59:13.340 --> 00:59:32.549 STFC-RAL-CR03 R61: And we had a very strong theory group at that time, and they supplied a lot of knowledge and information. We don't have that now, but we now know what they knew then, and we're trying to implement that into a large code, which goes into a supernova code, but we have to have general relativity 418 00:59:32.550 --> 00:59:36.100 STFC-RAL-CR03 R61: Fluid dynamics, the neutrino physics, etc. there as well. 419 00:59:36.100 --> 00:59:45.690 STFC-RAL-CR03 R61: So that's… that's a big problem now, and we don't have the resources that you have to solve some of these problems. Yes, yeah. 420 00:59:45.720 --> 00:59:49.189 STFC-RAL-CR03 R61: So that I'm… the Corsack model. 421 00:59:49.390 --> 00:59:53.639 STFC-RAL-CR03 R61: sort of… it tries to fix these things, I guess it's a… 422 00:59:53.960 --> 01:00:02.509 STFC-RAL-CR03 R61: they have these collaborations, which are loosely defined. So there's one in theoretical politics, there's one in numerical Relativity. 423 01:00:02.950 --> 01:00:11.599 STFC-RAL-CR03 R61: anyone can be part of those. It's not a… there's a PI, but they gather people, and then the community decides this. 424 01:00:12.080 --> 01:00:18.770 STFC-RAL-CR03 R61: spend the time doing things. There's a way there to either propose a new community, or to, 425 01:00:18.900 --> 01:00:23.250 STFC-RAL-CR03 R61: Join an existing one, and influencing them, say, this is an important thing that we need to then… 426 01:00:23.810 --> 01:00:29.969 STFC-RAL-CR03 R61: get into these people in there. That's the only… sanctioned method, I know. 427 01:00:30.630 --> 01:00:47.540 STFC-RAL-CR03 R61: Having studentships, joint studentships could help. Yes. That would be a very strong incentive to get more money coming in, say, from either SDFC, EBSRC, or wherever, and then having a joint studentship between departments. 428 01:00:48.530 --> 01:00:50.970 STFC-RAL-CR03 R61: Is there any mechanism that we can do that? 429 01:00:52.520 --> 01:01:10.210 STFC-RAL-CR03 R61: it gets really hard if you need to get another entity that is the university, right? Yes, you have to… that's a problem. Yes. There's not a problem getting a university involved. They're desperate for students. Yes, but they then want time with them, and then… Yeah, well… 430 01:01:10.270 --> 01:01:13.739 STFC-RAL-CR03 R61: If you've got the same problem that they're interested, then it's joining. 431 01:01:14.130 --> 01:01:16.099 STFC-RAL-CR03 R61: So we, they're all… 432 01:01:16.990 --> 01:01:32.920 STFC-RAL-CR03 R61: it doesn't help with particle physics, because there's a logic center signatures, so Robert is one of these people, he's doing a PhD with the University of Bath. So, S with diamond, and with the aristocrat, so there's a three-party problem there. 433 01:01:32.920 --> 01:01:36.940 STFC-RAL-CR03 R61: Jens was alluding to that, you have to have a universal work. 434 01:01:36.940 --> 01:01:39.470 STFC-RAL-CR03 R61: But it comes with challenges. Yes. 435 01:01:39.470 --> 01:01:58.840 STFC-RAL-CR03 R61: Yeah. But if everyone's on the same page, they want to get through the same job, it can work. Yeah. So you're, you're at, Europe, you're at Bath, but you're working here all the time, is that right? But you go to Bath? You're working here for a year? I'm working here, sort of, as a placement, sort of year, but it's not actually a year out of my PhD. Right. 436 01:01:59.100 --> 01:02:01.020 STFC-RAL-CR03 R61: So, can't accept me there. 437 01:02:01.240 --> 01:02:02.010 STFC-RAL-CR03 R61: Bye. 438 01:02:02.550 --> 01:02:14.690 STFC-RAL-CR03 R61: But where that money comes from the ELC, where the money does come from for the particles… So we need a lot of these have all the money in that. 439 01:02:15.430 --> 01:02:17.550 STFC-RAL-CR03 R61: Sorry, I have to ask Fortin. 440 01:02:18.570 --> 01:02:22.549 STFC-RAL-CR03 R61: Yes, please elaborate. 441 01:02:23.450 --> 01:02:36.120 STFC-RAL-CR03 R61: Yeah, I mean, both hands are very good line to solve the kind of problems that we solve. Yes. Does it make it challenging to move to new hardware? As Stuart said earlier, the… 442 01:02:36.120 --> 01:02:44.120 STFC-RAL-CR03 R61: So, you also spoke about NVIDIA versus AMD hardware. Do you use Open at all, or… 443 01:02:44.430 --> 01:02:46.440 STFC-RAL-CR03 R61: OpenCL, what's the thing? 444 01:02:46.680 --> 01:02:51.130 STFC-RAL-CR03 R61: There were backend libraries that were more inclusive, too. 445 01:02:52.160 --> 01:03:04.950 STFC-RAL-CR03 R61: Yeah, so the… so our GPU cord, GPU solver is a, like, a Frankenstein's monster underneath it, so it's foreground with C++ parts that bend this CUDA. 446 01:03:05.080 --> 01:03:14.520 STFC-RAL-CR03 R61: And, yeah, so it's very complicated to maintain and to write, but you needed the C++ template into the work properly. So, 447 01:03:14.670 --> 01:03:17.660 STFC-RAL-CR03 R61: Yeah, again, historically, it's Fortrone. 448 01:03:18.340 --> 01:03:37.350 STFC-RAL-CR03 R61: in the future, in 10 years' time, will the software be rewrite the Fortran? I don't think so. That's what we thought 10 years ago. But Fortran's still the most used software after 2, for example, so it's still quite widely used in practical tools and so on. 449 01:03:37.670 --> 01:03:39.440 STFC-RAL-CR03 R61: But, yeah, our philosophy is that 450 01:03:39.730 --> 01:03:55.020 STFC-RAL-CR03 R61: whatever it's written, there's interfaces to its main language that's called HSL, it's Python, with current reports in Python, there's MATLAB, and there's Julia, and there's C, and C++ that you can interface to. Doesn't matter what's in the backend. 451 01:03:56.590 --> 01:04:03.739 STFC-RAL-CR03 R61: we currently have developers that understand the portal, and we have… we could literally wrote a book on it, if we want new features. 452 01:04:03.880 --> 01:04:05.309 STFC-RAL-CR03 R61: Argue for that. 453 01:04:05.540 --> 01:04:09.710 STFC-RAL-CR03 R61: And then in 20 minutes, we might be implemented by compiles. 454 01:04:10.130 --> 01:04:11.019 STFC-RAL-CR03 R61: What's over. 455 01:04:11.140 --> 01:04:17.649 STFC-RAL-CR03 R61: Yeah, so modern full trend is a reasonable language, and it's very good for the kind of calculations we can do. 456 01:04:20.480 --> 01:04:25.519 STFC-RAL-CR03 R61: it has… There are issues where people don't want their full-time compiler. 457 01:04:25.780 --> 01:04:27.110 STFC-RAL-CR03 R61: But in the kitchen. 458 01:04:27.490 --> 01:04:29.139 STFC-RAL-CR03 R61: So, I will keep that. 459 01:04:31.440 --> 01:04:38.170 STFC-RAL-CR03 R61: There's a lot… a lot of people don't want to do a lot of programming, so what they're interested in is just having 460 01:04:38.340 --> 01:04:45.649 STFC-RAL-CR03 R61: the problem solved, and so they go to symbolic algebra, like Mathematica or something like that. So do you keep a… 461 01:04:45.980 --> 01:04:50.390 STFC-RAL-CR03 R61: A sort of suite of coins that you can… we can access. 462 01:04:51.250 --> 01:04:52.700 STFC-RAL-CR03 R61: What can you help us? 463 01:04:52.810 --> 01:04:59.269 STFC-RAL-CR03 R61: Yes. You know, navigate through the different symbolic algebra codes that are there. 464 01:05:03.720 --> 01:05:15.649 STFC-RAL-CR03 R61: Possibly, yes. I mean, yeah, we've used things like mathematical and can… But you had… I'll disturb that answer and throw one more in. You had a paper title there earlier about 465 01:05:16.090 --> 01:05:24.160 STFC-RAL-CR03 R61: I don't even remember what it was, but I think a lot of what you might be challenged by is how to communicate to people what you've delivered. 466 01:05:24.310 --> 01:05:34.240 STFC-RAL-CR03 R61: Like, how do you get people to engage with you if your pink or title says X? Exactly. Well, what it is is a Lego brick, and you need to communicate to people that it is, in fact, LEGO. 467 01:05:34.670 --> 01:05:35.550 STFC-RAL-CR03 R61: Thanks. 468 01:05:36.010 --> 01:05:38.960 STFC-RAL-CR03 R61: How… where do you go? How do you… 469 01:05:39.110 --> 01:05:41.130 STFC-RAL-CR03 R61: How do you get people to come to you? 470 01:05:41.480 --> 01:05:49.809 STFC-RAL-CR03 R61: Other than giving 7 months. Yes, but yeah, no, it is a challenge. It's very hard, and it's hard, 471 01:05:50.010 --> 01:05:53.390 STFC-RAL-CR03 R61: Yeah, we talked to… I think the thing that's… 472 01:05:55.130 --> 01:06:01.510 STFC-RAL-CR03 R61: the important thing, but what we do is it can plug in everywhere. If you talk to the public unions, people see, 473 01:06:02.340 --> 01:06:11.260 STFC-RAL-CR03 R61: what's the application that this enables? And we have lots of applications, and we do work with people, but it's a… but then this can also be applied to you, possibly apply to you. 474 01:06:12.290 --> 01:06:19.800 STFC-RAL-CR03 R61: So it… it's hard. We haven't solved it, we're trying different things. Getting folks like this helps. 475 01:06:23.170 --> 01:06:25.070 STFC-RAL-CR03 R61: I'm so coming in this… 476 01:06:25.310 --> 01:06:36.830 STFC-RAL-CR03 R61: in our structure at CFC, is there some entity which deals with all the licensing things, interfacing with the industry, taking care of the transactions and negotiations and stuff? 477 01:06:37.130 --> 01:06:38.559 STFC-RAL-CR03 R61: Who does… 478 01:06:40.190 --> 01:06:59.619 STFC-RAL-CR03 R61: So, within the Business and Innovations Directorate, there's an innovations team, and in there are the people that we deal with. What's the team called? Innovations. Innovations team. Yeah, and so Lizzie Bing is the, licensing… the IT manager. 479 01:06:59.690 --> 01:07:03.430 STFC-RAL-CR03 R61: And there's, adam Brochio, 480 01:07:04.510 --> 01:07:13.640 STFC-RAL-CR03 R61: Rapirez, who's the licensing manager and better, and then there's… there's also rehabby, business… 481 01:07:14.440 --> 01:07:20.829 STFC-RAL-CR03 R61: Yeah, there's one embedded within the team, but there's one that sits within innovations as well. 482 01:07:21.800 --> 01:07:26.309 STFC-RAL-CR03 R61: Yeah, and… I guess, I don't think it's live as well. 483 01:07:26.660 --> 01:07:32.569 STFC-RAL-CR03 R61: And that's the person who does negotiations with companies and stuff, but it's a bit of a… 484 01:07:32.730 --> 01:07:34.909 STFC-RAL-CR03 R61: Minefield, they are… 485 01:07:35.160 --> 01:07:40.769 STFC-RAL-CR03 R61: Yeah, I started with Paul Vernon yesterday, and they are hiring a software person for that team, because they… 486 01:07:41.230 --> 01:07:43.709 STFC-RAL-CR03 R61: They're very good at patents, and 487 01:07:44.030 --> 01:07:51.109 STFC-RAL-CR03 R61: widgets, but software is a different thing. They do not really understand software at the moment. They're very good, but they… 488 01:07:51.300 --> 01:08:02.430 STFC-RAL-CR03 R61: the kind of the software world. So someone is being added into that team with software, it's their core focus. I think that'd be… that'd be really good for every team, but have good… 489 01:08:02.550 --> 01:08:04.269 STFC-RAL-CR03 R61: These conversations and stuff. 490 01:08:05.050 --> 01:08:06.090 STFC-RAL-CR03 R61: I think so. 491 01:08:10.660 --> 01:08:11.530 STFC-RAL-CR03 R61: Okay. 492 01:08:12.650 --> 01:08:14.460 STFC-RAL-CR03 R61: Zoom? 493 01:08:15.430 --> 01:08:19.380 STFC-RAL-CR03 R61: I didn't see any question, but I think we had a very good discussion. 494 01:08:19.569 --> 01:08:29.130 STFC-RAL-CR03 R61: And you're nearby anyway, and now we are going for lunch. So, yes. Okay, let's close the seminar, and thank ours. 495 01:08:32.620 --> 01:08:34.599 STFC-RAL-CR03 R61: Thanks for talking, everyone.