Computational Fluid Dynamics Advances

🚀 𝐀𝐜𝐜𝐞𝐥𝐞𝐫𝐚𝐭𝐢𝐧𝐠 𝐂𝐅𝐃 - 𝐏𝐫𝐞𝐩𝐫𝐢𝐧𝐭 𝐀𝐥𝐞𝐫𝐭!  🚀 Excited to share our latest preprint: "Accelerating Computational Fluid Dynamics with Transported Memory Networks" 🧠🌊 In this work, we introduce Transported Memory Networks (TMNs) - inspired by Sepp Hochreiter’s LSTMs - as a novel approach to to capture the effect of unresolved scales in #FluidDynamics by means of #MachineLearning. Or more boldly: to 𝐚𝐜𝐜𝐞𝐥𝐞𝐫𝐚𝐭𝐞 𝐂𝐨𝐦𝐩𝐮𝐭𝐚t𝐢𝐨𝐧𝐚𝐥 𝐅𝐥𝐮𝐢𝐝 𝐃𝐲𝐧𝐚𝐦𝐢𝐜𝐬 (#CFD) 🔎The key #Insight? 🔍In CFD, CNNs do not just capture spatial relations but rather infer temporal dependencies. Essentially, convolutions collect information further downstream, thus effectively reaching back in time ⏳. Our approach takes a memory-based perspective where information is advected along the flow (Eulerian view). By leveraging LSTM-inspired architectures that explicitly incorporate gradient information, we demonstrate that the memory is de-facto transported. The network is trained via an autoregressive approach (solver-in-the-loop), ensuring robustness and better alignment with physics ⚙️. Why does this matter? ✅ More physically consistent #ML extended solvers 🏗️ ✅ Better suited for high-performance industrial CFD 🏭 ✅ A step closer to scalable, ML-based physics solvers 🔬 Of course, challenges remain, but we're making rapid progress toward ML-driven CFD for real-world applications! 💡 Big kudos to the main contributors Matthias Schulz and Gwendal Jouan, as well as Stefan Gavranovic and Daniel Berger for making this possible! 🎉 Check it out! Link in the comments! And of course always appreciate any feedback or thoughts. #MachineLearning #CFD #DeepLearning #FluidDynamics #AIForScience

💡 AI Meets Fluid Dynamics: A New Chapter Unfolds This week, DeepMind (in collaboration with top universities) revealed a breakthrough that could reshape how we understand the world’s most unruly systems: fluid flow. At its heart: the Navier–Stokes equations — the mathematical laws that govern everything from swirling smoke and ocean currents to airflow over wings and blood flow in our bodies. For over a century, these equations have challenged mathematicians and physicists alike. One of the biggest puzzles? Under extreme conditions, could they “blow up” — i.e. lead to infinite velocities or pressures (singularities)? That’s one of the Clay Institute’s Millennium Prize Problems. 🔍 What’s new DeepMind used Physics-Informed Neural Networks (PINNs) enhanced with mathematical insights to search for and uncover new families of unstable singularities in fluid equations. Remarkably, their results have been validated and shown to follow consistent patterns in key parameters (like the “blow-up” speed λ) as they explore higher orders of instability. Their approach blends deep math and AI, pushing toward computer-assisted proofs — where AI doesn’t just calculate but helps discover solutions that humans then verify. 🔄 Why it matters (beyond the math) Improved modeling of turbulence and fluid behavior = better weather forecasting, climate models, and disaster prediction. Engineering gains: aircraft, ships, wind turbines, energy systems — all depend on mastering fluid flows. In the grand scope: this is a proof point that AI can partner in the frontier of fundamental science, not just applied tasks. It accelerates the path toward the holy grail: fully resolving whether the Navier–Stokes equations always behave “nicely” or sometimes break. 🚀 A Glimpse into the Future — “Fluid AI”? Imagine if, one day, we could tame turbulence so precisely that fluid flows themselves become information media. In that world: Water, air, or other fluids act like “liquid circuits,” routing signals via controlled vortices or waves. Turbulence shifts from chaotic to a programmable substrate for computing. The difference between AI on silicon and AI in fluids might blur. While that’s still speculative, the DeepMind result is a stepping stone. We’re seeing the early promise of using AI to unlock deep laws—and eventually, to harness those laws in new forms of computing.

✈️ Would you trust engineers working with 1980s-era resolution to design a next-gen aircraft? In simulation, mesh cells are like pixels: the more you have, the more detail you capture. But until now, even advanced engineering teams were limited to coarse meshes—far from the fidelity needed to fully trust or iterate on complex systems. 🛠️ Simulations at industrial scale have long been the bottleneck in designing and operating complex systems—from aircraft to cars to energy infrastructure. For decades, most of the improvement has been coming from Moore's law - CPU speed improvements, rather than software (some of which still use Fortran, an insider admitted during one of my many reference calls...) Enters AI and an amazing team who've been obsessed with that problem for year. So far, most models have broken down beyond meshes with a few 100K cells on multiple GPUs. But yesterday, the Emmi AI team released AB-UPT, the first fluid dynamics model scaling beyond 150 million mesh cells, running on a single GPU, and delivering real-world physics accuracy. - 150 million mesh cells (no typo, that's 1000x more) - Real-world accuracy on a single GPU This is not just faster simulation—it’s AI-native simulation that finally bridges the gap between research demos and industrial-grade engineering. It means aircraft designers can explore concepts in minutes instead of weeks. It means complex systems can be simulated in real time as they operate. At Serena Data Ventures, together with Juliette, Matthieu, Floriane, Bertrand and Charline, we invest in foundational technologies that shift what’s possible in infrastructure and foundational software. Johannes, Dennis, Miks and their whole team are doing just that—and doing it fast. Hats off to all of them for this game-changing breakthrough! 📄 𝗥𝗲𝗮𝗱 𝘁𝗵𝗲 𝗽𝗮𝗽𝗲𝗿: https://lnkd.in/deX_zQWx 🤗 𝗧𝗿𝘆 𝘁𝗵𝗲 𝗺𝗼𝗱𝗲𝗹: https://lnkd.in/dmd-xtpR 💻 𝗔𝗰𝗰𝗲𝘀𝘀 𝘁𝗵𝗲 𝗰𝗼𝗱𝗲: https://lnkd.in/dZY6EW_P 🧪 𝗖𝗵𝗲𝗰𝗸 𝘁𝗵𝗲 𝗱𝗲𝗺𝗼: https://demo.emmi.ai/ #FoundationalAI #CFD

Can AI help solve Millennium Math problem in Fluid dynamics? We take an important step in this direction in our recent paper with physics-informed learning: High-Precision PINNs in Unbounded Domains: Application to Singularity Formulation in PDEs We develop a modular and a robust framework for training Physics-Informed Neural Networks (PINNs) to high precision on unbounded domains. This is a key step toward using AI in rigorous PDE analysis and computer-assisted proofs of singularities. Highlights: 1. Enforcing far-field asymptotics and non-degeneracy via hard constraints Leveraging self-scaled BFGS optimizers for stable convergence. 2. Achieving 4-digit improvement over state-of-the-art for 2D Boussinesq with fewer training steps. 3. Applicable to both smooth and non-smooth self-similar blowup profiles We believe this is a significant step toward integrating machine learning into the toolbox for tackling open problems in fluid dynamics and singularity formation (e.g. Navier-Stokes). https://lnkd.in/gnYqWUu5

🌟 Transforming Urban Wind Modeling with Physics-Informed AI 🌟 Traditional methods for predicting urban wind fields, like CFD simulations, are powerful but often time-consuming and computationally expensive. But what if we could predict wind dynamics faster without sacrificing accuracy? That's exactly what the authors Xuqiang Shao, Zhijian Liu, Siqi Zhang, Zijia Zhao, and Chenxing Hu achieved in their paper: "PIGNN-CFD: A Physics-Informed Graph Neural Network for Rapid Predicting Urban Wind Field Defined on Unstructured Mesh." 🔑 Key Highlights: 1️⃣ Faster Wind Field Prediction: The PIGNN-CFD model delivers predictions 1–2 orders of magnitude faster than traditional CFD simulations. 2️⃣ Physics-Informed Learning: By incorporating physical laws (RANS equations) directly into the training process, the model ensures accurate and reliable predictions. 3️⃣ Scalability: It generalizes well to large-scale urban environments, making it a promising tool for urban planning, air quality studies, and climate resilience efforts. 4️⃣ Real-World Validation:u The model leverages data from wind tunnel experiments (AIJ) and validated CFD simulations created using OpenFOAM. 💡 Why This Matters: Accurately modeling urban wind fields is critical for addressing environmental challenges like heat islands, air pollution, and pedestrian comfort. By integrating advanced graph neural networks and CFD-based data, this study paves the way for scalable and efficient urban climate solutions. 📈 Implications: The PIGNN-CFD framework offers a glimpse into the future of physics-informed machine learning, where simulation time is slashed, enabling rapid decision-making for urban designers, environmental scientists, and engineers. 💬 What are your thoughts on the application of machine learning in computational fluid dynamics? Let's discuss! #PhysicsInformedAI #CFD #MachineLearning #UrbanWindModeling #GraphNeuralNetworks #AIforClimateSolutions Link for the Paper https://lnkd.in/dD7Tbihp

Is a "ChatGPT for fluids" actually tractable? Could we build a model that is industrially useful and widely applicable? This is a question I’ve been debating for several years. Today, I’m pleased to share a new preprint, "Fluid Intelligence: A Forward Look on AI Foundation Models in Computational Fluid Dynamics,” written in collaboration with Johannes Brandstetter and Siddhartha Mishra It has been an absolute pleasure to work with Johannes and Sid on this and we spent many hours going back and forth on this paper, each bringing our own insights and opinions. I can honestly say that I personally learnt so much from this and that was part of the reason for wanting to publish a paper in the first place – to bridge the gap between the #ML and #CFD communities. I have seen immense interest (as well as some healthy scepticism too) about the potential of #AI in CFD and engineering more broadly, but applying the "scaling laws" seen in Large Language Models to the complex physics of fluids isn't straightforward. CFD is not language; it has unique constraints and high-dimensional inputs. In this paper, we have attempted to bridge the gap between the ML and CFD communities by: • Deconstructing the industrial CFD process to help ML researchers understand the complexity of the input space. • Proposing a theoretical scaling law that accounts for the costs of both data generation and model training. • Estimating the computational resources required to reach the "large-scale limit." We observe that while data generation is the main bottleneck today (which can be partially improved by incorporating high-fidelity transient data), the scaling laws reveal a "crossover point" where training costs dominate at the large-scale limit. The location of the cross-over point is highly dependant on the models ability to absorb data and has implications on the ultimate bottleneck in the development of a “ChatGPT for fluids”. We hope this provides a first estimate of "what it would really take" to build such a model. We certainly don’t have all the answers, and I’m sure we are wrong about several things, but we hope to motivate further study in this area. I look forward to hearing your thoughts! NVIDIA Emmi AI ETH Zürich

"Can Physics-Informed Neural Networks (PINNs) simulate 3D turbulence?" A question we've been asked repeatedly since developing the framework in 2017. After nearly a decade of progress, we now have a conclusive answer. For the first time, we demonstrate that PINNs can simulate fully developed turbulent flows in 2D and 3D by learning solutions directly from the Navier-Stokes equations without training data or computational grids. Key technical ingredients: -- PirateNet architecture for deep networks -- Causal training strategies -- Self-adaptive loss weighting -- SOAP optimizer for resolving gradient conflicts -- Time-marching with transfer learning Validation on challenging benchmarks: -- 2D Kolmogorov flow (Re = 10⁶) -- 3D Taylor-Green vortex (Re = 1,600) -- 3D turbulent channel flow (Re_τ = 550) Results accurately reproduce key turbulence statistics including energy spectra, enstrophy, and Reynolds stresses. This work demonstrates that PINNs can handle complex chaotic systems, though computational efficiency remains an important area for future improvements. It opens new possibilities for mesh-free modeling and hybrid simulation approaches in computational fluid dynamics. This research was led by the outstanding work of Sifan Wang at Yale University, with key contributions from Panos Stinis at Pacific Northwest National Laboratory and Shyam Sankaran at Penn Engineering. The work was supported by the DOE Advanced Scientific Computing Research program. 📄 Read the preprint here: https://lnkd.in/enGrJJKm #PINNs #CFD #Turbulence #ScientificComputing #MachineLearning #DOE #ASCR #AppliedMathematics

If you've ever pushed LES into real industrial geometry, you know the grid can make or break the entire workflow. We just published a new white paper on clipped Voronoi diagram meshing. It's a different way of thinking about grid generation: globally consistent, robust around complex surfaces, and designed for modern scale-resolving CFD. Whether you are exploring LES or looking to improve current setups, you'll find some useful insights. 👉 Here's the download link: https://ow.ly/BlN250XL9fx #meshing #computationalfluiddynamics #largeeddysimulation

Heterogeneities, tyranny of scales, and curse of dimensionality In most complex physical systems (and to some extent others like financial markets), competing non-linear forces lead to instabilities and emergent large scale structures. The overall dynamics are controlled by the large scale structures to a great extent but those large structures cannot be accessed directly as they are the result of the small scale interactions. In CFD (Computational Fluid Dynamics), we are stuck in the pursuit of conducting even higher resolving simulations to get to the ground truth by modeling all those small scale interactions that lead to the emergent large scale structures. We are creative in coming up with subgrid models to reduce the computational load to resolve absolutely what we have to and have correlations for the unresolved scales. There are other heuristics that we apply to things like turbulence and drag closures, etc. to make the problems computationally tractable. Having a taste of methods to unravel low dimensional manifolds (my work with Badri Velamur Asokan back in 2007) and also realizing that most systems we are trying to model lie in much lower dimensions (agent based modeling work) than than the millions or now billions of degrees we bring in through CFD, I presented concepts at various venues on how we can use AI/ML to break these vicious dependencies between heterogeneities, tyranny of scales, and curse of ever increasing mesh sizes. Now it is more imperative to revisit as the LLM boom brought us into our laps the neural networks based learning algorithms that give access to low dimensional latent spaces and possibly transfer learning from similar phenomena, hardware co-designed with software for deep learning on large datasets, and plethora of software and data analysis tools available to explore all aspects of heterogeneous structures in complex systems. Below is a short blog post describing my attempt to develop a Dynamic Heterogeneity-Resolving Drag Model (DHRDM) for coarse-grid simulations of FCC risers: https://lnkd.in/gunwY5jQ Interactive website for you to explore this approach is at: https://lnkd.in/grE97R9M Sankaran Sundaresan Madhava (Syam) Syamlal Hans Kuipers Olivier Simonin Raffaella Ocone Shankar Subramaniam Tingwen Li Wei Wang Wei Ge Sanjib Das Sharma, Ph.D Marc-Olivier Coppens David West