# AI for Science in Action: Using Today's Accelerated Discovery and What’s Next ## Abstract Key takeaways Hear real-world case studies across fields on how researchers and engineers apply AI to accelerate discovery—from tool selection to deployment in production workflows. AI surrogate models are compressing scientific simulation times from months on supercomputers to seconds, fundamentally changing the pace of discovery. Get a broader view of where AI for science is heading—understand how trends emerging in HPC and research communities are converging with industry, and what that means for the future of your work. ## AI Summary - The panel discussed the integration of AI in scientific simulations, particularly in astrophysics, computational fluid dynamics, and climate modeling, highlighting how AI can accelerate and enhance these processes. - The speakers shared their experiences and challenges with data bottlenecks, emphasizing the need for better data management and analysis tools to handle the vast amounts of data generated by high-fidelity simulations. - The importance of trust and validation in AI models was highlighted, with a focus on comparing surrogate models with high-fidelity simulations and experimental data to ensure accuracy and reliability. - The panelists discussed the potential of quantum computing to bridge the gap between real-world phenomena and simulations, particularly in complex systems like weather and nuclear reactors. - Open science and the sharing of data, models, and code were emphasized as crucial for advancing research and ensuring that AI models are transparent and reproducible. - The panel concluded with a discussion on the need for more responsive and scalable computational resources in Europe to support the training and development of large AI models, and the importance of collaboration across different scientific domains. ## Transcript **[00:00:05 – 00:00:08]** Good, thank you so much, and good afternoon, and **[00:00:08 – 00:00:10]** welcome to this panel. **[00:00:10 – 00:00:12]** I'm Giuseppe Fiamini from NVIDIA. **[00:00:12 – 00:00:15]** I'm a solution architect and a leader team in Europe doing **[00:00:15 – 00:00:18]** research collaboration projects with the leading academic **[00:00:18 – 00:00:22]** institutions across Europe, but also private industry. **[00:00:22 – 00:00:24]** So, why we are here? **[00:00:24 – 00:00:27]** The goal is to really share with you what AI for science **[00:00:27 – 00:00:32]** is and how AI can accelerate different, I mean, scientific **[00:00:32 – 00:00:34]** workloads and numerical simulation. **[00:00:34 – 00:00:40]** AI is not only powering LLMs, hopefully, but it's also helping **[00:00:40 – 00:00:44]** a lot of scientists across Europe do better science, okay? **[00:00:44 – 00:00:50]** And now we see how classic HPC numerical simulations are somehow **[00:00:50 – 00:00:52]** merging with AI workloads, okay? **[00:00:52 – 00:00:55]** You can use AI to do pre-processing, post-processing. **[00:00:55 – 00:00:58]** You can use agents to run very complex workflow **[00:00:59 – 00:01:00]** and do optimization. **[00:01:00 – 00:01:04]** You can also replace some numerical solvers with the **[00:01:05 – 00:01:06]** AI components, okay? **[00:01:06 – 00:01:09]** We have no slides today. **[00:01:09 – 00:01:15]** We have four scientists from across different institutions **[00:01:15 – 00:01:21]** in Europe that will help me going through three relevant topics. **[00:01:22 – 00:01:23]** So, what works today? **[00:01:23 – 00:01:25]** What is currently working today? **[00:01:25 – 00:01:29]** How to make these results be trustworthy, and what's **[00:01:29 – 00:01:31]** coming next, okay? **[00:01:31 – 00:01:34]** And at the end, I hope there will be time for some questions **[00:01:35 – 00:01:38]** that you might have for our egregious panelists. **[00:01:38 – 00:01:40]** So I'll start from Marit Korpilang. **[00:01:40 – 00:01:42]** Welcome to this panel. **[00:01:42 – 00:01:45]** So Marit is professor at the University of Aalto, **[00:01:45 – 00:01:50]** University in Finland, and she's leading the astroinformatics **[00:01:50 – 00:01:53]** group in the HPC lab. **[00:01:53 – 00:01:57]** She studied the magnetic activity of the sun and turbulent plasmas. **[00:01:57 – 00:02:00]** She pioneered us informatics in Finland. **[00:02:00 – 00:02:04]** She also works on energy-efficient methods for research. **[00:02:04 – 00:02:07]** Marit, I have just one sentence, just to break the ice. **[00:02:07 – 00:02:15]** So, what is the scientific problem you most want AI to help solve? **[00:02:15 – 00:02:19]** We are working on developing more affordable modeling tools **[00:02:19 – 00:02:22]** for solar convection and its magnetism to understand the **[00:02:23 – 00:02:25]** roots of space weather. **[00:02:25 – 00:02:27]** Thank you, Mari. **[00:02:27 – 00:02:29]** I'll move to Oriol. **[00:02:29 – 00:02:35]** So Oriol Lehmkuhl is from the Barcelona **[00:02:35 – 00:02:37]** Supercomputing Center in Spain. **[00:02:37 – 00:02:40]** Oriol leads the large-scale computational fluid **[00:02:40 – 00:02:43]** dynamics groups at BSC. **[00:02:43 – 00:02:46]** He works on turbulence, multiphase flows, and aerodynamics at **[00:02:46 – 00:02:48]** a very large scale. **[00:02:48 – 00:02:52]** And his group developed the SOD2D, a simulation code built **[00:02:52 – 00:02:54]** to accelerate computing. **[00:02:54 – 00:02:57]** And SOD2D has also been used for flow control with **[00:02:57 – 00:02:59]** reinforcement learning. **[00:02:59 – 00:03:01]** So Oriol, also one sentence for you. **[00:03:01 – 00:03:05]** What is the problem that you would like AI to help solve? **[00:03:05 – 00:03:09]** So I will focus on my domain, on aeronautics, where we **[00:03:09 – 00:03:15]** have huge problems always when we go to reality, from **[00:03:15 – 00:03:18]** simulation to reality, passing through wind tunnel. **[00:03:18 – 00:03:21]** We have huge difference always, and this comes because we **[00:03:21 – 00:03:25]** have a lot of uncertainties on boundary conditions, on **[00:03:25 – 00:03:29]** physical phenomena that happen for the multi-physics aspects **[00:03:29 – 00:03:32]** of those problems, and because the geometries, they evolve, **[00:03:32 – 00:03:34]** they change through operation. **[00:03:34 – 00:03:38]** I would like that IA help us to close this gap. **[00:03:38 – 00:03:41]** Okay, thank you. I'll move to Chris Payne. **[00:03:41 – 00:03:45]** Christopher Payne, professor at Imperial College in London. **[00:03:45 – 00:03:48]** Chris leads the applied modeling and computational group at **[00:03:48 – 00:03:52]** Imperial with around more than 70 scientists. **[00:03:52 – 00:03:56]** It's one of the largest research group at Imperial. **[00:03:56 – 00:03:59]** And his group builds models for ocean, the atmosphere **[00:03:59 – 00:04:02]** and climate, and for complex flows. **[00:04:03 – 00:04:08]** Recently, Chris has been bringing AI into high fidelity simulation, **[00:04:08 – 00:04:14]** and, for example, his group developed the AI for PDs, **[00:04:14 – 00:04:20]** which is a physics solver that is working like a neural network, and **[00:04:21 – 00:04:26]** he has been publishing more than 300 journal papers into this field. **[00:04:26 – 00:04:28]** Chris, I have the same question for you. **[00:04:28 – 00:04:33]** What do you think AI should be, what do you want AI to solve? **[00:04:33 – 00:04:36]** Which kind of problems do you think AI should be solving, **[00:04:36 – 00:04:38]** help solving at least? **[00:04:38 – 00:04:43]** So I'd like AI to help provide the greatest possible resolution, **[00:04:43 – 00:04:47]** numerical resolution, for things like climate, weather, **[00:04:47 – 00:04:53]** direct numerical simulation of the air flows across the entire planet, **[00:04:53 – 00:05:00]** and to be able to bridge the scales from molecular scale to the **[00:05:00 – 00:05:05]** continuum and small scale continuum up to the planetary scale. **[00:05:05 – 00:05:07]** Thank you, Chris. **[00:05:07 – 00:05:13]** And finally, sitting next to Chris, we have Sadiq Jaffer **[00:05:13 – 00:05:15]** from University of Cambridge. **[00:05:15 – 00:05:18]** So Sadiq is assistant research professor in computer science **[00:05:18 – 00:05:23]** and the Bernstein Planetary Computing Fellow, uses AI **[00:05:23 – 00:05:27]** machine learning for nature and climate, is a co-author of Tessera. **[00:05:28 – 00:05:32]** So Tessera is an open foundation model that learns from satellite **[00:05:32 – 00:05:36]** images of the earth over time. **[00:05:36 – 00:05:40]** And Tessera was also present at a CVPR conference this year. **[00:05:40 – 00:05:44]** So, I mean, this is also interesting because this is **[00:05:44 – 00:05:48]** the first time that I see, I mean, papers from, let's say, classic **[00:05:48 – 00:05:55]** HPC domains being presented at very important AI conferences like CVPR. **[00:05:55 – 00:05:58]** Sadiq, what is your, I mean, your touch on this? **[00:05:58 – 00:06:02]** I mean, on this sentence, so which problem would you **[00:06:02 – 00:06:04]** like AI to help solving? **[00:06:04 – 00:06:05]** Sure, thank you. **[00:06:05 – 00:06:10]** So, our planet is changing ever more quickly, and what **[00:06:10 – 00:06:15]** I would like to see is models that can help us understand **[00:06:15 – 00:06:18]** what's changing, and then what are at least the short-term **[00:06:18 – 00:06:20]** implications of that change. **[00:06:20 – 00:06:22]** And we have all of the sensors for doing this. **[00:06:22 – 00:06:24]** We have satellites, and we have drones, and we have weather **[00:06:24 – 00:06:28]** balloons, and we have lots of data that can tell us what's changing. **[00:06:28 – 00:06:33]** And we have causality in academic literature and in supply chains **[00:06:33 – 00:06:34]** and that kind of thing. **[00:06:34 – 00:06:37]** What we lack is the AI models to connect that all together. **[00:06:37 – 00:06:38]** And that's what I would like to see. **[00:06:38 – 00:06:40]** Thank you, Sadiq. **[00:06:41 – 00:06:43]** Good, I think we can start with covering the first topic. **[00:06:43 – 00:06:44]** So what's the current status? **[00:06:44 – 00:06:46]** And we'll start with from you, Marit. **[00:06:46 – 00:06:49]** So Marit, let's start from the plasma, so which is a **[00:06:49 – 00:06:51]** field that you, well known. **[00:06:51 – 00:06:53]** So it's well known for you. **[00:06:53 – 00:06:58]** So simulation of the sun are probably one of the most intensive **[00:06:59 – 00:07:02]** from the computer, compute intensive, okay, in science. **[00:07:02 – 00:07:07]** And how do you think AI and HPC can change the way simulations **[00:07:07 – 00:07:12]** for model stellar, Dynamo, or solar activity can be conducted? **[00:07:12 – 00:07:14]** And what do you think it's, I mean, which kind **[00:07:14 – 00:07:17]** of applications do you see next? **[00:07:18 – 00:07:30]** Right, so... We are now heading the kind of GPU-assisted revolution **[00:07:30 – 00:07:33]** of computational astrophysics. **[00:07:33 – 00:07:38]** But in a couple of years, we will again face another performance war. **[00:07:38 – 00:07:42]** Even if we accelerate the simulations with GPUs, we cannot **[00:07:42 – 00:07:49]** really reach the parameter regimes of the real star, like our sun. **[00:07:49 – 00:07:54]** So what can we then do if the so-called classical **[00:07:54 – 00:07:57]** computing hits a wall? **[00:07:57 – 00:08:03]** The fact that we are aware of is that these high-fidelity **[00:08:03 – 00:08:05]** simulations that we are currently performing, they **[00:08:05 – 00:08:09]** are producing a huge amount of... high-fidelity data **[00:08:09 – 00:08:14]** that we can use for understanding these physical systems. **[00:08:14 – 00:08:16]** And this is the only way to go forward. **[00:08:16 – 00:08:20]** So to use this high-fidelity data in understanding the **[00:08:20 – 00:08:22]** physical process is better. **[00:08:22 – 00:08:27]** And there we can leverage machine learning and AI tools a lot. **[00:08:27 – 00:08:30]** And this has not been done, especially in the magnetized **[00:08:30 – 00:08:31]** plasma simulation physics. **[00:08:31 – 00:08:37]** We are really lagging behind the hydrodynamics, like CFD **[00:08:37 – 00:08:40]** or climate, because we are dealing with magnetized plasmas, **[00:08:40 – 00:08:44]** which have very complex phenomena occurring in them, meaning **[00:08:44 – 00:08:48]** that we have inverse cascades where large-scale phenomena **[00:08:48 – 00:08:55]** arise from turbulent flows, and this is like we are taking **[00:08:55 – 00:09:00]** the first steps towards using ML and AI to understand the data. **[00:09:00 – 00:09:05]** And even that is not without challenges, because the data **[00:09:05 – 00:09:10]** that we are at the moment producing from the GPU simulations **[00:09:10 – 00:09:15]** is reaching petabyte scale in terms of disk space. **[00:09:15 – 00:09:19]** So we have a serious data bottleneck, where to store **[00:09:19 – 00:09:21]** it and how to deal with it. **[00:09:21 – 00:09:26]** So this old-fashioned kind of workflows where you do the data **[00:09:26 – 00:09:32]** analysis in post-processing, these are all dead in our community. **[00:09:32 – 00:09:36]** So what we have to do is to analyze the data as it comes. **[00:09:36 – 00:09:40]** So we have to stream it from the memory of the simulation **[00:09:40 – 00:09:46]** itself to some machine learning tool and try to learn on the fly. **[00:09:47 – 00:09:54]** This is the current paradigm, and the most prominent problem **[00:09:54 – 00:09:57]** at the moment is the data analysis bottleneck. **[00:09:57 – 00:10:02]** It's faster to produce the data than analyze it, and **[00:10:02 – 00:10:07]** we are struggling with it, but we hope to overcome this bottleneck **[00:10:07 – 00:10:12]** and then leverage this ML and AI tools to produce more affordable **[00:10:13 – 00:10:16]** simulation tools, which we call large eddy simulations, which **[00:10:16 – 00:10:19]** are known for climate and CFD. **[00:10:19 – 00:10:23]** I think these people at least know what I mean. **[00:10:23 – 00:10:27]** But to describe the physics at the smaller scales, to **[00:10:27 – 00:10:32]** be able to build more affordable simulation tools for the whole **[00:10:33 – 00:10:35]** system at large scale. **[00:10:35 – 00:10:37]** Thank you, Marit. So you really raise a very important **[00:10:37 – 00:10:39]** aspect of your work. **[00:10:39 – 00:10:43]** So how to really, I mean, analyze large-scale data? **[00:10:43 – 00:10:49]** I mean, as long as you collect them. **[00:10:49 – 00:10:54]** So you want to collect data and immediately analyze them, okay? **[00:10:54 – 00:10:58]** Large-scale, from data coming from, I guess, telescopes, coming from **[00:10:58 – 00:11:02]** sensors or from satellite images. **[00:11:03 – 00:11:04]** Thank you so much. **[00:11:04 – 00:11:08]** Just a curiosity, but why does **[00:11:09 – 00:11:13]** space weather matter for us humans? **[00:11:13 – 00:11:16]** I mean, maybe this is a question that was not part **[00:11:16 – 00:11:19]** of our outline, but just came to the top of my mind. **[00:11:19 – 00:11:24]** Can you elaborate more on that, Marit? **[00:11:24 – 00:11:25]** What is space weather? **[00:11:25 – 00:11:32]** Yeah, why, I mean, space weather matter for us, for humans? **[00:11:32 – 00:11:38]** Right, yeah, so space weather sounds kind of innocent, but **[00:11:38 – 00:11:41]** it's not for our modern society. **[00:11:41 – 00:11:46]** So we are based on digital society nowadays, and the sun **[00:11:46 – 00:11:52]** is a very dangerous electromagnetic object that can... ignite **[00:11:52 – 00:11:58]** space storms that directly affect our infrastructure here on Earth **[00:11:58 – 00:12:01]** and also communications in space. **[00:12:01 – 00:12:02]** Thank you so much. **[00:12:02 – 00:12:06]** Let's go back to the heart, Oriol, from space to the heart. **[00:12:07 – 00:12:08]** So, and turbulence. **[00:12:08 – 00:12:11]** So at BSC you run turbulence simulations at very large **[00:12:11 – 00:12:15]** scale using your SO2D tool. **[00:12:16 – 00:12:23]** So, where does AI fit in studying turbulence and large-scale **[00:12:23 – 00:12:27]** simulation and the daily work that you conduct at the BSC? **[00:12:28 – 00:12:33]** It's quite similar to what was described just before. **[00:12:33 – 00:12:40]** We did begin two, three years ago just with the ability to compute **[00:12:40 – 00:12:43]** very huge cases thanks to GPU. **[00:12:43 – 00:12:47]** And this has generated huge data sets, and the first need **[00:12:47 – 00:12:49]** was to compress these data sets. **[00:12:49 – 00:12:54]** And NVIDIA has been very useful to that, in order to not need **[00:12:54 – 00:12:58]** to download the whole database, to be able to compress it with several **[00:12:58 – 00:13:02]** physical-based methods, but they use machine learning as well, that **[00:13:02 – 00:13:06]** we can run online, and then as a result of our simulations, we share **[00:13:06 – 00:13:12]** more of the model, and not the big raw data that is petabyte level. **[00:13:12 – 00:13:16]** But then after, we were using as well IA to do reinforced **[00:13:16 – 00:13:22]** learning to optimise flow conditions, to optimise geometries **[00:13:22 – 00:13:26]** of the geometries that we are solving, and we realised **[00:13:26 – 00:13:30]** that it is still very expensive to go optimising with high fidelity. **[00:13:31 – 00:13:35]** And then those compressed type of technologies we extended to **[00:13:35 – 00:13:42]** be able to have surrogates that we can introduce on our optimizations. **[00:13:42 – 00:13:44]** And now we have been working a lot on that, on how to **[00:13:44 – 00:13:48]** build these surrogates that understand the system. **[00:13:48 – 00:13:52]** We know how to do this very well on a steady state type of conditions. **[00:13:52 – 00:13:55]** With transient conditions, it's an open question that **[00:13:55 – 00:13:59]** my community, at least on our level, we are discussing. **[00:13:59 – 00:14:03]** And it's what we believe that we need to close now. **[00:14:03 – 00:14:10]** And right now the challenge is to tackle transient and **[00:14:10 – 00:14:15]** to begin to introduce agentic workflows that in these last **[00:14:15 – 00:14:16]** six months it has explored. **[00:14:16 – 00:14:20]** We are seeing that they are very useful on this optimization loop. **[00:14:20 – 00:14:23]** They are very useful as well to help us to improve **[00:14:24 – 00:14:25]** codes and methods. **[00:14:25 – 00:14:29]** But now we are discovering how to integrate this on our **[00:14:29 – 00:14:30]** type of workflows. **[00:14:30 – 00:14:32]** Thank you, Oriol. **[00:14:32 – 00:14:33]** Very, very interesting. **[00:14:33 – 00:14:37]** So, I mean, variational autoencoders have always been **[00:14:37 – 00:14:40]** used also to do data compression. **[00:14:40 – 00:14:45]** It's like mapping the insights of our datasets, the features of **[00:14:45 – 00:14:48]** our datasets, into the latent space of this variational autoencoder. **[00:14:48 – 00:14:51]** And if I got it right, this is something that you are **[00:14:51 – 00:14:56]** exploring to reduce the size of the data sets that you deal with. **[00:14:56 – 00:15:01]** And you've been also using a term which is known to us, but maybe **[00:15:01 – 00:15:03]** the whole community doesn't know. **[00:15:03 – 00:15:04]** It's surrogate. **[00:15:04 – 00:15:08]** So can you explain a few words what a surrogate model is? **[00:15:08 – 00:15:12]** Yeah, a surrogate model for us, usually in our community, **[00:15:12 – 00:15:18]** was we have the full model that needs a lot of freedoms, billions **[00:15:18 – 00:15:21]** of points, where we know velocity. **[00:15:21 – 00:15:27]** And there are some techniques based on SVD type of linear **[00:15:27 – 00:15:33]** analysis where you can reduce other models on some... **[00:15:35 – 00:15:41]** On some vectors, some vector directions that can reproduce **[00:15:41 – 00:15:43]** this type of model. **[00:15:43 – 00:15:46]** This, without encoders and other techniques, transformers, **[00:15:46 – 00:15:50]** etc., we can do non-linear instead of linear mathematics, **[00:15:50 – 00:15:55]** and then we can have a small model with fewer degrees of freedom. **[00:15:56 – 00:15:58]** The problem is that sometimes now we lose the physics, and **[00:15:59 – 00:16:02]** we are trying to understand and to have some methods of doing **[00:16:02 – 00:16:04]** this without losing the physics. **[00:16:04 – 00:16:09]** The traditional methods, all these vectors are orthogonal, **[00:16:09 – 00:16:13]** so we know that they don't interact between them, etc. **[00:16:14 – 00:16:17]** And this reduced model then can be used in order to assess **[00:16:17 – 00:16:20]** conditions that were not on the original simulation. **[00:16:20 – 00:16:25]** It's kind of a foundational model, but for a specific domain. **[00:16:25 – 00:16:28]** Definitely. Thank you so much, Oriol. **[00:16:28 – 00:16:31]** So, Chris, let's stay with the simulation. **[00:16:31 – 00:16:37]** Your group uses AI surrogates for many multi-physics problems **[00:16:37 – 00:16:43]** and applications, from CFD to particle physics to air quality. **[00:16:43 – 00:16:46]** Can you share, let's say, concrete results where an **[00:16:46 – 00:16:51]** AI model replaced, or really can be replaced to use, I mean, instead **[00:16:51 – 00:16:54]** of a classic numerical simulation? **[00:16:54 – 00:16:57]** Okay, yeah, happy to do that. **[00:16:58 – 00:17:01]** So probably the most spectacular result that we've produced **[00:17:01 – 00:17:08]** so far is a foundational surrogate model that was trained on **[00:17:08 – 00:17:13]** a very small part of London, South Kensington area, and then it was **[00:17:13 – 00:17:17]** applied to all of Greater London, everything within the M25, about **[00:17:17 – 00:17:20]** 40 kilometers by 40 kilometers. **[00:17:20 – 00:17:27]** So that model had half a meter resolution and corresponding **[00:17:27 – 00:17:32]** time resolution, so it's resolving the turbulent fluctuations. **[00:17:32 – 00:17:37]** Yet the model that we generated was able to run a thousand **[00:17:37 – 00:17:40]** times faster than reality. **[00:17:41 – 00:17:45]** So that's very important for us to use that as a basis **[00:17:45 – 00:17:51]** for helping to manage and design our cities, for example, **[00:17:51 – 00:17:57]** bringing in data from smart cities, indoors, outdoors, managing **[00:17:57 – 00:18:00]** our ventilation systems indoors. **[00:18:00 – 00:18:03]** One could even imagine with that kind of resolution, or perhaps **[00:18:03 – 00:18:10]** a bit more resolution, even sort of resolving some of the The air flows **[00:18:10 – 00:18:12]** within this room, for example. **[00:18:12 – 00:18:17]** So we presented those results at the Prime Minister's office **[00:18:17 – 00:18:24]** in India because of their concerns about the pollution **[00:18:24 – 00:18:33]** within Delhi, and with the idea of using these tools to help design **[00:18:33 – 00:18:38]** and manage Delhi, and perhaps other cities across the world. **[00:18:38 – 00:18:44]** The second aspect, I think, is probably a neurophysics **[00:18:44 – 00:18:47]** model, so I only ask for one. **[00:18:47 – 00:18:51]** I think I need to probably present another example of **[00:18:51 – 00:18:55]** what's called a neurophysics model, wherein we've injected the physics **[00:18:55 – 00:18:56]** directly into the neural network. **[00:18:56 – 00:19:01]** There's no training, but you specify the weights analytically **[00:19:01 – 00:19:07]** so that you essentially solve the discretized PDEs. **[00:19:07 – 00:19:10]** So that was important, and we applied that to the same **[00:19:10 – 00:19:11]** Greater London area. **[00:19:11 – 00:19:15]** That was important because it used another concept that **[00:19:15 – 00:19:20]** we're trying to develop that we refer to as intelligent resolution. **[00:19:20 – 00:19:23]** So you put the resolution in the models where it's needed, and you **[00:19:23 – 00:19:25]** take it out where it's not needed. **[00:19:26 – 00:19:30]** So that had a resolution of five centimetres, but across **[00:19:30 – 00:19:32]** all of Greater London again. **[00:19:32 – 00:19:35]** It wasn't nearly as fast, but it had that five **[00:19:35 – 00:19:37]** centimetre resolution. **[00:19:37 – 00:19:43]** So I believe that these methods are showing transformative advantages. **[00:19:43 – 00:19:48]** The neurophysics models are essentially equivalent to **[00:19:48 – 00:19:51]** classical CFD or classical models, but they have many **[00:19:51 – 00:19:56]** of the advantages of the AI models, for example, differentiability, **[00:19:56 – 00:19:59]** bringing in data, and so on. **[00:19:59 – 00:20:00]** Thank you very much, Chris. **[00:20:00 – 00:20:04]** What do you think, I mean, the trade-offs are between simulation **[00:20:04 – 00:20:07]** and AI and surrogates models? **[00:20:07 – 00:20:10]** The trade-off, I mean, when to go for a numerical simulation, **[00:20:10 – 00:20:13]** classic numerical simulation, when to, let's say, try to **[00:20:13 – 00:20:15]** train a surrogate model? **[00:20:15 – 00:20:19]** Yeah, the trade-off. **[00:20:19 – 00:20:26]** So, I think it depends on where you, where you, what **[00:20:26 – 00:20:27]** you want to achieve. **[00:20:27 – 00:20:31]** If you absolutely want to have the highest level of confidence, **[00:20:31 – 00:20:33]** I think you have to insert the PD. **[00:20:33 – 00:20:35]** The PDEs, the governing equations in there. **[00:20:36 – 00:20:40]** If you want to extend the resolution, perhaps increase **[00:20:40 – 00:20:43]** confidence because you got more resolution, then you could start **[00:20:43 – 00:20:45]** adding in the machine learning. **[00:20:45 – 00:20:50]** And I believe the best approach is a combination of both, **[00:20:51 – 00:20:56]** where you inject the physics in, and then you resolve subsequent **[00:20:56 – 00:21:01]** scale processes with some of the AI surrogate light models. **[00:21:01 – 00:21:02]** So, yeah. **[00:21:02 – 00:21:04]** Thank you so much, Chris. **[00:21:04 – 00:21:09]** So, Sadiq, we've been heard a lot about simulations, no? **[00:21:09 – 00:21:12]** Your work starts from observation instead, no? **[00:21:12 – 00:21:16]** You apply AI to climate and nature, and from Tessera's open **[00:21:16 – 00:21:18]** global embeddings to conservation. **[00:21:18 – 00:21:23]** What does accelerated discovery look like in your own group work? **[00:21:23 – 00:21:24]** Thank you, Giuseppe. **[00:21:24 – 00:21:28]** So just very quickly, Tessera is an Earth observation foundation model. **[00:21:28 – 00:21:33]** So we take essentially a year's worth of optical and radar imagery **[00:21:33 – 00:21:38]** from satellites, and we compress that down into a 128-dimension **[00:21:38 – 00:21:42]** embedding or fingerprint for every 10 meters of land NVIDIA, AI. **[00:21:42 – 00:21:47]** And you can use that instead of the raw satellite data. **[00:21:47 – 00:21:49]** And if you do, you get two big advantages. **[00:21:49 – 00:21:55]** First is you get a huge reduction in data volume, about 40 times. **[00:21:55 – 00:21:57]** But the second one is that you get a large increase in **[00:21:57 – 00:21:58]** label efficiency. **[00:21:58 – 00:22:01]** It is far more efficient to do these kind of tasks on top of **[00:22:01 – 00:22:05]** these compact latent, pre-trained latent representations than **[00:22:05 – 00:22:07]** it is to do it on the raw data. **[00:22:08 – 00:22:13]** Now half of my group are actually plant scientists and ecologists, **[00:22:13 – 00:22:17]** and so for us kind of accelerated development is that we can **[00:22:17 – 00:22:23]** produce maps and models at a scale that we just, we couldn't before. **[00:22:23 – 00:22:26]** And that's partly because it's significantly less data **[00:22:26 – 00:22:29]** volume, but also it's because sometimes we just cannot get **[00:22:29 – 00:22:32]** more ground truth data for the kind of task we're trying to do. **[00:22:32 – 00:22:36]** And there are things like habitat mapping or land use, **[00:22:36 – 00:22:40]** land cover, change detection, **[00:22:40 – 00:22:43]** crop classification, crop health monitoring, those kinds **[00:22:43 – 00:22:47]** of tasks that in the past would never have a hope of **[00:22:47 – 00:22:51]** doing a region scale map or model to understand how these things **[00:22:51 – 00:22:53]** are changing, and now we can. **[00:22:53 – 00:22:57]** So for us, that's what we see as accelerated discovery. **[00:22:57 – 00:23:01]** Thank you so much. I mean, who is currently using Tessera today? **[00:23:01 – 00:23:03]** How many users do you have? **[00:23:03 – 00:23:05]** I mean, what's your roadmap for Tessera? **[00:23:05 – 00:23:07]** So we think it's about a hundred and... **[00:23:07 – 00:23:09]** When we last counted, there's about a hundred and eighty different **[00:23:09 – 00:23:12]** organizations that are using it. **[00:23:12 – 00:23:14]** From kind of industry, we know we have a lot **[00:23:14 – 00:23:16]** of government users as well. **[00:23:16 – 00:23:21]** It's essentially replacing parts of, you know, manual **[00:23:21 – 00:23:23]** remote sensing pipelines. **[00:23:23 – 00:23:27]** They're based on the raw data, and they've done a side-by-side **[00:23:27 – 00:23:30]** with some of these geospatial embeddings and realized they **[00:23:30 – 00:23:33]** can get similar or better performance by replacing those kind **[00:23:33 – 00:23:38]** of hand-engineered, domain-specific kind of pipelines, so that's **[00:23:38 – 00:23:39]** mainly where they're coming from. **[00:23:39 – 00:23:41]** Thank you so much. **[00:23:42 – 00:23:46]** We have heard that with AI you can do faster science, okay, **[00:23:46 – 00:23:49]** but fast is not enough in science. **[00:23:49 – 00:23:52]** So AI can accelerate workloads, but fast is not enough. **[00:23:52 – 00:23:56]** We also need trust; we need to mechanisms to trust and **[00:23:56 – 00:24:00]** validate the results that we get from AI, from AI models, no? **[00:24:00 – 00:24:03]** And we'll start from you, Amarit, and then jump to Sadiq, because **[00:24:03 – 00:24:11]** you all work with observation data that you collect from telescope or **[00:24:11 – 00:24:13]** from satellite images or sensors. **[00:24:13 – 00:24:19]** So about openness, you have open weights, open embeddings, **[00:24:19 – 00:24:23]** open code, but what does open really mean in science? **[00:24:23 – 00:24:28]** What does open give to the science today? **[00:24:28 – 00:24:33]** Yeah, for us, open code, and if we were able to share open **[00:24:33 – 00:24:35]** models, that gives a lot. **[00:24:35 – 00:24:42]** It's a high-gain, not-so-much-effort business. **[00:24:43 – 00:24:45]** But what is now hindering our research is really **[00:24:45 – 00:24:46]** the data bottleneck. **[00:24:46 – 00:24:52]** So at the moment, when we are not yet in the state that **[00:24:52 – 00:24:56]** we can actually only share the surrogate models, because **[00:24:56 – 00:25:00]** we don't have them, we are just developing them, so now we **[00:25:00 – 00:25:03]** are at the moment in the situation where we have to learn from **[00:25:03 – 00:25:05]** the data that we have produced. **[00:25:05 – 00:25:09]** And this is now reaching like hundreds of terabytes scale. **[00:25:09 – 00:25:15]** So if we talk about open data, it's kind of a joke, what we do now. **[00:25:15 – 00:25:19]** Because if you think about a terabyte scale, one single **[00:25:19 – 00:25:22]** system state, it cannot be shared with anyone. **[00:25:22 – 00:25:28]** It's like it will be situated in the HP system where it's produced. **[00:25:28 – 00:25:32]** It cannot be moved, it cannot be downloaded, it cannot be analyzed. **[00:25:32 – 00:25:35]** It's accessible only to a few people. **[00:25:35 – 00:25:38]** So, even if we would make the link to this data open, **[00:25:38 – 00:25:41]** nobody could use it. **[00:25:42 – 00:25:48]** So, now it's a bit like we don't have the correct and proper **[00:25:48 – 00:25:52]** processes to analyze this big data. **[00:25:52 – 00:25:55]** We would like to make it accessible to everyone, but we do not **[00:25:55 – 00:25:59]** have the capability of doing that, because nobody can download **[00:25:59 – 00:26:02]** it, nobody can actually analyze it. **[00:26:02 – 00:26:04]** The only thing what we can do is to produce more and **[00:26:04 – 00:26:09]** more data, because computation is cheap, but storage and **[00:26:09 – 00:26:13]** bandwidth are not anymore. **[00:26:13 – 00:26:15]** Glad to hear that computation is cheap. **[00:26:16 – 00:26:20]** At the moment, it's a bit like we would also need to **[00:26:20 – 00:26:24]** address this data bottleneck for certain domains, maybe **[00:26:24 – 00:26:29]** not yours, but like certain domains where we do not yet have **[00:26:29 – 00:26:31]** the surrogate models available. **[00:26:31 – 00:26:39]** So that kind of would need a change of mindset and also politics **[00:26:39 – 00:26:45]** to make more people to access this data and take use of it. **[00:26:45 – 00:26:46]** Thank you so much, Merlin. **[00:26:46 – 00:26:49]** Jumping to Sadiq, what do you think Open is giving to **[00:26:50 – 00:26:51]** science in your own field? **[00:26:51 – 00:26:55]** Yes, so for us, Open is about trust. **[00:26:55 – 00:27:02]** So we have users that use Tessera to produce data that then informs **[00:27:02 – 00:27:05]** policy or maybe a directly making policy, and those policy decisions **[00:27:05 – 00:27:08]** can have implications for decades. **[00:27:08 – 00:27:14]** And for us, it's crucial to have provenance of the data that's gone **[00:27:14 – 00:27:19]** into these AI models and also to be able to ensure that, for decades **[00:27:19 – 00:27:23]** to come, people can take those models and they can do inference **[00:27:23 – 00:27:26]** with them, they can update them, they can manipulate them, even if **[00:27:26 – 00:27:31]** the lab or the organization that's produced them no longer exists. **[00:27:31 – 00:27:34]** I don't think it's right to be basing these decisions **[00:27:34 – 00:27:38]** on, you know, black box proprietary models where, you know, the **[00:27:38 – 00:27:41]** kitchen sink of data may have gone into it, so we have no idea **[00:27:41 – 00:27:44]** what we can rely on as outputs, and where there is no guarantee **[00:27:44 – 00:27:46]** that, you know, they're going to be updated for years to come. **[00:27:46 – 00:27:51]** So for us, it's It's about trust, but a cost comes with **[00:27:51 – 00:27:56]** that openness, so as we're saying, data is expensive. **[00:27:56 – 00:27:59]** The Tessera embeddings globally are about two petabytes of **[00:27:59 – 00:28:03]** data, and that has to be made available and high bandwidth, **[00:28:03 – 00:28:08]** and we're only able to do that thanks to support from industrial **[00:28:08 – 00:28:11]** collaborators, which we're very lucky to have, but that's not **[00:28:11 – 00:28:13]** always the case for other groups. **[00:28:13 – 00:28:15]** I think that's something that needs to be, if we want to do more **[00:28:15 – 00:28:18]** open science, that's something that has to be funded and resourced, **[00:28:18 – 00:28:23]** so you're doing the science; let's make sure that the outputs of it, **[00:28:23 – 00:28:26]** we also resource to make available. **[00:28:26 – 00:28:27]** Thank you so much, Salik. **[00:28:28 – 00:28:33]** So, going back to simulations now and trust, how can you **[00:28:33 – 00:28:37]** really trust the results of the surrogate models that you generate? **[00:28:37 – 00:28:41]** When we talk about, for instance, conservation laws, uncertainty **[00:28:41 – 00:28:44]** quantification, or comparison with high-fidelity models, **[00:28:44 – 00:28:46]** simulations, sorry. **[00:28:46 – 00:28:49]** How can you really, I mean, first of all, how do you **[00:28:49 – 00:28:52]** validate the results, and how you can trust them? **[00:28:52 – 00:28:53]** Please, Oriol. **[00:28:53 – 00:28:56]** Yeah, so to me it's very clear. **[00:28:56 – 00:28:59]** Always we have this process where we compare with the high fidelity. **[00:28:59 – 00:29:02]** If it's a surrogate, we need to compare. **[00:29:02 – 00:29:07]** The first principle preservation, I believe, it's part of the **[00:29:07 – 00:29:11]** training because it's part of what we try to embed to the **[00:29:11 – 00:29:14]** architecture of the model, right? **[00:29:14 – 00:29:19]** But then, always we need to compare with the high fidelity, and **[00:29:19 – 00:29:22]** in our case the high fidelity needs to be compared with experiments. **[00:29:22 – 00:29:27]** So we have this other loop to ensure that we can follow the **[00:29:27 – 00:29:29]** real experiments on the surrogates. **[00:29:29 – 00:29:34]** And that the good thing is that we can assess on the uncertainty, **[00:29:34 – 00:29:37]** because it's one of the things that with machine learning we can do. **[00:29:37 – 00:29:41]** We can assess uncertainty and understand where the surrogate **[00:29:41 – 00:29:45]** has more uncertainty and inject more high fidelity data there. **[00:29:45 – 00:29:50]** So I believe this loop of all the time looking for the uncertainty, **[00:29:50 – 00:29:54]** understanding where there are more problems, inject more data, etc. **[00:29:54 – 00:29:57]** Until you have a surrogate that you understand the errors and how **[00:29:57 – 00:30:03]** the errors are managed, it's a very important loop to do on our field. **[00:30:03 – 00:30:05]** Yeah, for instance, I mean, it's interesting that we don't **[00:30:05 – 00:30:08]** have a sort of, this is something that we've been, we would **[00:30:08 – 00:30:11]** like to work a bit in NVIDIA, we don't have a leaderboard for **[00:30:11 – 00:30:15]** AI for Science open weights model. **[00:30:15 – 00:30:18]** So there are a lot of leaderboards available for classical LMs **[00:30:18 – 00:30:22]** with benchmarks, but we don't have a leaderboard, or at **[00:30:22 – 00:30:26]** least a set of benchmarks for... We have huge discussions. **[00:30:26 – 00:30:30]** Actually, right now on Fluid Mechanics, there is an initiative **[00:30:30 – 00:30:33]** to have a Fluid Bench. **[00:30:33 – 00:30:35]** And this is going to come in the next month. **[00:30:35 – 00:30:39]** Right now, we are finalizing to understand which cases **[00:30:39 – 00:30:44]** we want to To release, how we want to build the leaderboards and **[00:30:44 – 00:30:47]** to do all the type of assessment. **[00:30:47 – 00:30:48]** And this is going to come. **[00:30:48 – 00:30:53]** It's a necessary step if we want to... No, definitely, definitely. **[00:30:53 – 00:30:55]** Thank you so much. **[00:30:55 – 00:30:56]** Chris, still on trust and validation. **[00:30:56 – 00:30:58]** What's your take on this? **[00:30:59 – 00:31:04]** So I think the classical methods of comparing the models with **[00:31:04 – 00:31:09]** experiments and other models, maybe the classical, previous **[00:31:09 – 00:31:14]** generation models, is a very good approach, but I think there are **[00:31:14 – 00:31:23]** sort of new opportunities presented with AI, for example, bringing **[00:31:23 – 00:31:28]** in data explicitly into the models and then quantifying remaining **[00:31:28 – 00:31:30]** uncertainties within the models. **[00:31:30 – 00:31:37]** And I think that should be seen as a... almost a standard approach **[00:31:37 – 00:31:42]** for verification and validation of models, and I'd like to see **[00:31:42 – 00:31:51]** sort of an automated approach for models using that kind of thinking. **[00:31:51 – 00:31:54]** Thank you so much. **[00:31:54 – 00:31:57]** European Commission, but I mean nations in Europe, Europe **[00:31:57 – 00:32:01]** in general is doing a lot to provide them in public **[00:32:01 – 00:32:04]** computational resources, so there are many systems available, **[00:32:04 – 00:32:06]** and I know, Chris, that you are using a lot in Imperial, I mean **[00:32:06 – 00:32:12]** Eisenbar AI system, you've been using Mare Nostrum and Leonardo, **[00:32:12 – 00:32:17]** Marit, you as well, Sadiq, so what do you think Europe should be doing **[00:32:17 – 00:32:18]** to help you with this respect? **[00:32:18 – 00:32:22]** Do you expect, I mean, you have been mentioning availability **[00:32:22 – 00:32:25]** or large-scale storage to store all the data that you **[00:32:25 – 00:32:29]** collect from the sky, but, I mean, what do you think Europe **[00:32:29 – 00:32:36]** should be doing that it's not doing now, in terms of support, help? **[00:32:36 – 00:32:38]** We are all belonging to academic, public academic **[00:32:38 – 00:32:40]** institution, you know? **[00:32:40 – 00:32:43]** And European Union and European countries are doing a lot **[00:32:43 – 00:32:48]** to support your research and also to help you be more competitive **[00:32:48 – 00:32:52]** with respect to large, let's say, America corporate, for instance. **[00:32:52 – 00:32:56]** So, OpenAI, Androbic, NVIDIA itself are doing a lot of research. **[00:32:56 – 00:32:58]** But, I mean, we don't want to compete. **[00:32:58 – 00:33:01]** I mean, we are here to support you, but, I mean, what do **[00:33:02 – 00:33:05]** you think European Union, but European countries should **[00:33:05 – 00:33:10]** be doing to help you be more competitive in that respect? **[00:33:10 – 00:33:12]** Marit, if you want to start. **[00:33:12 – 00:33:13]** Right. **[00:33:13 – 00:33:19]** Well, I think the European landscape has been driven **[00:33:19 – 00:33:25]** quite strongly with constructing infrastructure, but not building **[00:33:25 – 00:33:31]** any sort of, you know, human resources infrastructure around it. **[00:33:31 – 00:33:34]** So we are still slightly suffering from the lack **[00:33:34 – 00:33:39]** of skills, lack of skills and lack of skills to develop codes for **[00:33:39 – 00:33:41]** the infrastructure is what we have. **[00:33:42 – 00:33:47]** And then the understanding of what the capabilities are **[00:33:47 – 00:33:54]** when you really develop software that can scale up in these **[00:33:54 – 00:33:56]** exascale infrastructures, what kind of data masses you **[00:33:56 – 00:34:01]** have, and the implications for the society to utilize **[00:34:01 – 00:34:04]** this data is really lacking behind. **[00:34:04 – 00:34:06]** So we have this marvelous infrastructure; we should **[00:34:06 – 00:34:10]** now develop skills to use it properly, I think. **[00:34:10 – 00:34:12]** Thank you so much. **[00:34:12 – 00:34:17]** Yeah, so I believe, as you say, so the industrial strength **[00:34:17 – 00:34:20]** on the States is very big, no? **[00:34:20 – 00:34:26]** On this AI machine learning area, but Europe has an opportunity **[00:34:26 – 00:34:31]** because we have this federation of Auto-HPC machines. **[00:34:31 – 00:34:35]** And I believe we can do a step forward to have a federation **[00:34:35 – 00:34:42]** on EF4Science and to have really a network to do EF4Science, **[00:34:42 – 00:34:48]** to embrace open access and open science on these processes. **[00:34:48 – 00:34:52]** And maybe we will not have that big infrastructure, but **[00:34:52 – 00:34:56]** all the ecosystem, the different universities accessing to **[00:34:56 – 00:35:00]** this further learning, and to be open, that everything **[00:35:01 – 00:35:04]** is clear, how the researchers have been engaging to the agents, **[00:35:04 – 00:35:08]** have been doing their research, and after their data being open **[00:35:08 – 00:35:12]** and accessible, if we do that, then we have an opportunity to remain **[00:35:12 – 00:35:14]** competitive on the scientific side. **[00:35:14 – 00:35:16]** So being more open and then... **[00:35:16 – 00:35:18]** I think if provision could be provided for people and **[00:35:18 – 00:35:24]** agents maybe to help develop the NVIDIA, DRIVEOS, HPC, AI. **[00:35:45 – 00:35:49]** In light of the, you know, the agents like Claude that **[00:35:49 – 00:35:55]** are being used to develop these models, I think that's **[00:35:55 – 00:36:02]** going to present a whole new suite of challenges and opportunities. **[00:36:02 – 00:36:06]** I'd also like maybe **[00:36:07 – 00:36:15]** So maybe agents and people on the application side of the models **[00:36:15 – 00:36:21]** as well, very different disciplines to analyze the results, interpret **[00:36:21 – 00:36:27]** them, and so on and so forth. **[00:36:27 – 00:36:30]** But I also think that there's potentially a need to bring **[00:36:30 – 00:36:33]** data into these models, stream NVIDIA, DRIVEOS, HPC, AI. **[00:36:33 – 00:36:35]** NVIDIA, DRIVEOS, HPC, AI. **[00:36:35 – 00:36:37]** NVIDIA, DRIVEOS, HPC, AI. NVIDIA, DriveOS, HPC, AI. **[00:36:37 – 00:36:42]** A high-performance computing system so that one could provide **[00:36:42 – 00:36:49]** a weather forecasting system or a city forecasting system that could **[00:36:49 – 00:36:57]** be used to help manage every facet of our, of our cities, for example. **[00:36:57 – 00:37:01]** So maybe combining the high-performance computing **[00:37:01 – 00:37:08]** with operational demands might be a possible sort **[00:37:08 – 00:37:12]** of way forward as well. **[00:37:12 – 00:37:16]** So openness kills cooperation. **[00:37:16 – 00:37:25]** And yeah, I mean, I think we should never forget that numerical **[00:37:25 – 00:37:28]** simulations were born in Europe. **[00:37:28 – 00:37:35]** So there is a strong focus on applications in Europe. **[00:37:35 – 00:37:39]** Maybe we are not developing hardware, we are not developing **[00:37:39 – 00:37:47]** chips for the ARM experience in Cambridge, UK, but in Europe **[00:37:47 – 00:37:53]** there is a large set of expertise on applications across multiple **[00:37:53 – 00:37:56]** scientific domains. **[00:37:56 – 00:37:58]** Sadiq, what's your take here? **[00:37:59 – 00:38:03]** So I think part of at least my experience in the UK is, **[00:38:03 – 00:38:05]** and this is from the training foundation model's point of **[00:38:05 – 00:38:08]** view, where if you're training a new type of foundation model, **[00:38:08 – 00:38:12]** for one, you don't know how and whether it will scale. **[00:38:12 – 00:38:15]** And the way that we allocate compute, at least in the UK, **[00:38:15 – 00:38:18]** is kind of more of a traditional HPC model, where I've got **[00:38:18 – 00:38:21]** something that I know work, I want to run it at this size, **[00:38:21 – 00:38:23]** and I want this amount of compute. **[00:38:23 – 00:38:26]** When you're trying to train a foundation model and you actually **[00:38:26 – 00:38:30]** don't know whether it's going to scale, you start small, you get **[00:38:30 – 00:38:32]** some positive results, and then you want to scale up your computation. **[00:38:32 – 00:38:35]** You now might wait eight months to get another allocation **[00:38:35 – 00:38:37]** to be able to do that, at which point the people you have **[00:38:37 – 00:38:40]** with all the skills are kind of spinning their wheels and not able **[00:38:40 – 00:38:42]** to do something, or they've left. **[00:38:42 – 00:38:45]** So I think we need a bit more of a responsive, especially if **[00:38:45 – 00:38:47]** we want people to train foundation models, we need more of a **[00:38:47 – 00:38:52]** responsive way of having these kind of gated scaling towards bigger **[00:38:52 – 00:38:55]** and bigger models where you can show that you can meet those, there **[00:38:55 – 00:38:59]** is compute available kind of almost immediately to go to the next step. **[00:38:59 – 00:39:02]** Otherwise, it's very, very hard to train these in an **[00:39:02 – 00:39:03]** academic institution in the UK. **[00:39:03 – 00:39:04]** Thank you so much, Sadiq. **[00:39:04 – 00:39:07]** So, I would love really to continue this conversation, but we are **[00:39:07 – 00:39:09]** reaching the top of this session. **[00:39:09 – 00:39:12]** But if you don't mind, I would like to take questions from **[00:39:12 – 00:39:16]** the audience, if any, maybe one, two questions for our panelists. **[00:39:16 – 00:39:18]** Is there any question from you? **[00:39:18 – 00:39:20]** Any interest? Any... **[00:39:20 – 00:39:23]** Please. Do we have a mic? **[00:39:25 – 00:39:26]** We have, let's say, two minutes. **[00:39:26 – 00:39:30]** Otherwise, we had the last session, so we are lucky. **[00:39:30 – 00:39:31]** We have a more... Hi, guys. **[00:39:32 – 00:39:35]** My name is Luca. I've been loving the panel so far. **[00:39:35 – 00:39:37]** One thing that I've really been realizing here is that **[00:39:37 – 00:39:40]** in the workflows that you guys do work with in this **[00:39:40 – 00:39:44]** very physical, real world, where the simulations are failing, where **[00:39:44 – 00:39:47]** do you think that quantum computing is going to bridge that gap **[00:39:47 – 00:39:49]** between real world and simulation? **[00:39:49 – 00:39:52]** And do you think that it will get closer to being able to **[00:39:52 – 00:39:55]** be that real thing that will be able to leverage the metaphysics **[00:39:55 – 00:39:58]** that really do get in the way of having those real world **[00:39:58 – 00:39:59]** results from the simulation? **[00:39:59 – 00:40:02]** Thank you so much. We are opening another dimension of the problem. **[00:40:02 – 00:40:04]** Who wants to take this quiz question? **[00:40:04 – 00:40:06]** Please, Chris. Thank you. **[00:40:06 – 00:40:07]** Thanks for the question. **[00:40:07 – 00:40:08]** Very good question. **[00:40:08 – 00:40:12]** So from my perspective, I think the possibility of introducing **[00:40:12 – 00:40:19]** quantum computing into physics model is just so exciting. **[00:40:19 – 00:40:25]** The possibility of modeling perhaps the air flows and **[00:40:25 – 00:40:29]** the weather across the entire... A planet with direct numerical **[00:40:29 – 00:40:33]** simulation is just mind-blowing for me, you know, millimeter **[00:40:33 – 00:40:36]** or less kind of resolution. **[00:40:36 – 00:40:40]** That is the only possibility currently. **[00:40:40 – 00:40:44]** But, you know, modeling things like a nuclear reactor from the **[00:40:44 – 00:40:48]** molecular scale upwards suddenly also becomes a possibility. **[00:40:48 – 00:40:51]** I think we can think very differently about the sorts **[00:40:51 – 00:40:54]** of problems that we could solve. **[00:40:54 – 00:40:59]** I'd like to participate in that, if that's possible. **[00:41:01 – 00:41:03]** Thank you, Chris. **[00:41:03 – 00:41:04]** Merita, I think you said something. **[00:41:04 – 00:41:05]** Yeah, very quickly. **[00:41:05 – 00:41:11]** I hope, of course, that we already solved this issue with the **[00:41:11 – 00:41:14]** surrogates that we developed from the high-fidelity simulations that **[00:41:14 – 00:41:16]** we can do with classical computing. **[00:41:16 – 00:41:22]** But if we don't, then at least in my research, it will mean **[00:41:22 – 00:41:27]** to design completely the modeling gear again, because quantum **[00:41:27 – 00:41:32]** computing is a bit different beast than what we are doing now. **[00:41:32 – 00:41:35]** But it's possible, but we hope to solve it before we **[00:41:35 – 00:41:37]** have to go there. **[00:41:37 – 00:41:39]** Let's see what we can do. Yeah, step by step. **[00:41:39 – 00:41:42]** One last final question, please, **[00:41:42 – 00:41:46]** before Jensen kills me **[00:41:49 – 00:41:53]** And I get laid off. **[00:41:54 – 00:41:57]** Thank you. On the issue of open data and open massive **[00:41:58 – 00:42:01]** data, do the panel consider it enough, I hope so, because this **[00:42:01 – 00:42:06]** is what I do, to provide the input scripts as the open data so that **[00:42:06 – 00:42:11]** others can regenerate your data, if not now, then in five or ten years? **[00:42:12 – 00:42:15]** I will add input scripts and prompts. **[00:42:15 – 00:42:17]** Who wants to take this? **[00:42:17 – 00:42:18]** Please, Marit. **[00:42:18 – 00:42:21]** Yeah, we already do that, but this is not enough. **[00:42:21 – 00:42:25]** I mean, we are sharing everything we do, like scripts, also **[00:42:25 – 00:42:27]** for data analysis. **[00:42:27 – 00:42:30]** But this is not enough, because the data is sitting in our **[00:42:30 – 00:42:33]** HPC system somewhere in Finland, and nobody else has access. **[00:42:34 – 00:42:36]** So, how to provide the access? **[00:42:37 – 00:42:40]** I mean, not only for the input scripts that we have, but maybe **[00:42:40 – 00:42:44]** somebody else has more intelligent scripts, more intelligent **[00:42:44 – 00:42:49]** machine learning pipelines to implement on our data, so it should **[00:42:49 – 00:42:54]** be bi-directional, not only us providing something, but the public **[00:42:54 – 00:42:59]** at large providing something to work on the data that we have, and **[00:42:59 – 00:43:02]** this is not enabled at the moment; we don't have a process for that. **[00:43:02 – 00:43:04]** I guess I'm getting at, you said yourself, computers **[00:43:04 – 00:43:07]** are cheap, storage and bandwidth are expensive. **[00:43:07 – 00:43:10]** If other people can generate your data themselves, is that enough? **[00:43:10 – 00:43:14]** And then you can work on their variation. **[00:43:14 – 00:43:18]** Maybe somebody else's data, but not ours, because we are dealing with **[00:43:18 – 00:43:22]** the magnetic Dynamo process, which takes like, even with GPUs, months **[00:43:22 – 00:43:25]** to mature in a supercomputer. **[00:43:25 – 00:43:28]** If it would be something that you can calculate in a five-minute **[00:43:28 – 00:43:32]** thingy on GPUs, that would be the case, but in my realm, **[00:43:32 – 00:43:33]** it's not the case. **[00:43:33 – 00:43:38]** Maybe you can... Yeah, so we do provide the input scripts **[00:43:38 – 00:43:38]** to be able to regenerate. **[00:43:38 – 00:43:42]** It's just to do that requires moving about 100 petabytes **[00:43:42 – 00:43:46]** of data, and it's about 3 million pounds to redo the computation. **[00:43:46 – 00:43:51]** So the reason why we include the pre-generated outputs is because if **[00:43:51 – 00:43:54]** we only gave you the input scripts, like, it would suddenly price **[00:43:54 – 00:43:58]** out nearly everybody can be able to actually reproduce that work. **[00:43:58 – 00:44:01]** So as an example, 10 years ago I was doing simulations **[00:44:01 – 00:44:04]** producing terabytes of data, couldn't share that, made **[00:44:04 – 00:44:05]** the input scripts available. **[00:44:05 – 00:44:07]** You can run that on a desktop in an hour now. **[00:44:07 – 00:44:10]** You don't need HPC. So in that respect, the input **[00:44:11 – 00:44:11]** scripts were enough. **[00:44:11 – 00:44:15]** It is now shared. People can build on that. **[00:44:15 – 00:44:16]** Perfect. **[00:44:16 – 00:44:18]** Thank you so much for coming. **[00:44:18 – 00:44:20]** We've reached the end of this session. **[00:44:20 – 00:44:23]** Thank you. **[00:44:23 – 00:44:28]** And please join me; give me a big applause to our panelists. **[00:44:28 – 00:44:30]** Thank you, Meric, thank you, Aureole, thank you, Chris, **[00:44:30 – 00:44:31]** thank you, Sadiq. **[00:44:31 – 00:44:32]** Thank you so much. **[00:44:32 – 00:44:34]** And enjoy the rest of GTC.