The Neil Ashton Podcast

Why leave NVIDIA to join Mistral?In this episode, Neil Ashton explains his move to Mistral as VP of Computational Engineering, where he will work alongside Johannes Brandstetter on AI for engineering, physics foundation models and large language models.Neil reflects on the career decisions that took him from Manchester and Formula 1 to Oxford, NASA collaborations, AWS and NVIDIA. He shares what he learned from moving between academia and industry, how open datasets became central to his work, and why the chance to keep learning and build something new made Mistral an exciting next step.For students, researchers and engineers considering a change, the episode also offers a personal perspective on developing a specialism, keeping an academic mindset in industry, and staying open to opportunities without needing to plan your entire career in advance.Topics include:Getting a first break in Formula 1Switching between academia and industryLearning HPC, cloud computing and product engineeringMachine learning, open datasets and scientific collaborationNVIDIA's hardware, software and partnership ecosystemWhy Mistral, and working with Johannes BrandstetterCareer advice: expertise, continual learning and embracing changePodcast archive: https://neilashton.co.uk/podcasts/Chapters:00:00 Podcast intro00:39 From NVIDIA to Mistral01:46 My new role: physics AI and LLMs03:18 Starting out in engineering and motorsport06:26 A PhD and finding a specialism08:38 Getting a first break in Formula 109:52 Returning to academia11:57 Oxford, HPC and NASA15:21 Consulting, Formula 1 and the connection to AWS19:39 Leaving Oxford and joining AWS21:06 Product engineering, ML and open datasets23:29 What I learned at NVIDIA26:39 Why Mistral felt like the next step29:51 Career advice: stay open to change31:28 Becoming known for a specialism32:41 Podcast plans and closing thoughtsPaper mentioned:Fluid Intelligence: A Forward Look on AI Foundation Models in Computational Fluid DynamicsNeil Ashton, Johannes Brandstetter and Siddhartha Mishrahttps://arxiv.org/abs/2511.20455These are personal reflections on my own career and experience.

Should you still study engineering when AI can already write code, analyse data and automate parts of an engineer's job?In this solo episode, Neil Ashton gives his view on engineering education and careers in the age of AI. His answer is yes—but the skill set is changing. Neil explains why engineering fundamentals still matter, where AI can act as an enabler, what students and early-career engineers should learn now, and why soft skills, projects and internships may become even more important.Topics include:- Why demand for engineers is likely to remain strong- The engineering tasks most likely to change- AI as an enabler for coding, CAD, CAE and automation- Why domain knowledge is still essential for checking AI's work- What practical AI fluency means beyond using a chat interface- Advice for undergraduate, postgraduate and PhD students- How projects, internships and soft skills can help you stand outPodcast archive: https://neilashton.co.uk/podcasts/Chapters:00:00 Podcast intro00:39 The career question in the age of AI03:20 Why engineering demand is still growing04:43 Which engineering tasks AI will change05:21 AI as an engineering enabler09:18 Why fundamentals and domain expertise still matter11:59 AI fluency and the hiring market16:41 Advice for students and researchers20:05 What engineers should study now22:11 Standing out: soft skills, projects and internships25:41 Is engineering still worth it?Resource mentioned:- World Economic Forum, Future of Jobs Report 2025 — Skills outlook: https://www.weforum.org/publications/the-future-of-jobs-report-2025/in-full/3-skills-outlook/Please note that this episode expresses my personal opinion and does not represent the views of NVIDIA.

Hydrogen combustion, high-fidelity CFD and the future of aircraft propulsion are the focus of this conversation with Dr. Daniel Mira, Head of the Propulsion Technologies Group at the Barcelona Supercomputing Center. Neil and Dani discuss why reacting flows are so difficult to simulate, how hydrogen changes combustion and aircraft design, the limits of RANS, LES and DNS, GPU-native solvers, coding agents and AI surrogate models.Full episode, corrected transcript and resources:https://neilashton.co.uk/podcasts/s4-e6-daniel-mira-on-hydrogen-combustion-modelling-and-future-propulsion/TopicsWhy reacting flows are so computationally difficultHydrogen versus hydrocarbon combustionWhen hydrogen could reach commercial aviationHow engines and aircraft must be redesignedIndustrial trust in high-fidelity combustion CFDRANS, LES and DNS for reacting flowsChemistry, load balancing and computational costWall modelling in combustion LESGPU acceleration and solver redesignCoding agents for scientific softwareAI surrogate models and digital engineering workflowsSelected resourcesDaniel Mira and the Propulsion Technologies Grouphttps://ptg.bsc.es/?p=44Propulsion Technologies Group — research lineshttps://ptg.bsc.es/research-lines/BSC — Combustion researchhttps://www.bsc.es/research-development/research-areas/engineering-simulations/combustionCenter of Excellence in Combustion (CoEC)https://coec-project.eu/High-fidelity simulations of the mixing and combustion of a technically premixed hydrogen flamehttps://upcommons.upc.edu/entities/publication/08a27c10-cb13-4357-a3ab-8e9ec1d706ccChapters00:00 Podcast intro00:39 Introducing Daniel Mira03:00 Conversation begins04:55 Why combustion CFD is so hard10:23 Daniel’s path into hydrogen and jet-engine combustion12:48 Hydrogen versus hydrocarbon combustion17:58 Industrial adoption of hydrogen20:54 Gas turbines, aviation and fuel infrastructure25:35 How jet engines must change30:43 Redesigning the whole aircraft34:46 What will trigger commercial adoption?37:27 Why aerospace projects take a decade42:14 RANS, LES and DNS for reacting flows44:31 Replacing expensive tests with high-fidelity CFD46:01 The biggest accuracy gaps in combustion LES49:26 Where the computational cost goes52:06 Chemistry, species and source-term bottlenecks55:35 Wall modelling in combustion LES59:49 GPUs, algorithms and solver redesign01:08:52 Can coding agents accelerate combustion CFD?01:12:27 AI surrogate models for combustion01:24:20 Closing thoughts

Differentiable physics, neural emulators and foundation models for PDEs are the focus of this conversation with Professor Nils Thuerey, head of the Physics-based Simulation group at TUM. Neil and Nils discuss PhiFlow, PICT, Tadpole, scalable 3D transformers, online synthetic data, open datasets, world models and agents that call physics simulators.Full episode, corrected transcript and resources:https://neilashton.co.uk/podcasts/s4-e5-prof-nils-thuerey-on-differentiable-physics-and-foundation-models/TopicsDifferentiable physics and physics-based deep learningPhiFlow and differentiable simulation across ML frameworksWhen neural emulators can outperform their training dataFoundation models for PDEs and synthetic online trainingScalable 3D transformers and high-resolution simulationsLES, temporal data and correlated CFD datasetsOpen-source tools, startups and physics-aware world modelsAI agents that call physics simulatorsPapersNeural Emulator Superiority: When Machine Learning for PDEs Surpasses its Training Datahttps://arxiv.org/abs/2510.23111Tadpole: Autoencoders as Foundation Models for 3D PDEs with Online Learninghttps://arxiv.org/abs/2605.15284P3D: Scalable Neural Surrogates for High-Resolution 3D Physics Simulations with Global Contexthttps://arxiv.org/abs/2509.10186PICT — A Differentiable, GPU-Accelerated Multi-Block PISO Solver for Simulation-Coupled Learning Tasks in Fluid Dynamicshttps://arxiv.org/abs/2505.16992PhiFlow: Differentiable Simulations for PyTorch, TensorFlow and JAXhttps://proceedings.mlr.press/v235/holl24a.htmlPhysics-based Deep Learninghttps://arxiv.org/abs/2109.05237Learning to Control PDEs with Differentiable Physicshttps://arxiv.org/abs/2001.07457Solver-in-the-Loop: Learning from Differentiable Physics to Interact with Iterative PDE-Solvershttps://arxiv.org/abs/2007.00016tempoGAN: A Temporally Coherent, Volumetric GAN for Super-resolution Fluid Flowhttps://arxiv.org/abs/1801.09710Deep Learning Methods for Reynolds-Averaged Navier-Stokes Simulations of Airfoil Flowshttps://arxiv.org/abs/1810.08217WeatherBench: A Benchmark Dataset for Data-Driven Weather Forecastinghttps://arxiv.org/abs/2002.00469SuperWing: A Comprehensive Transonic Wing Dataset for Data-Driven Aerodynamic Designhttps://arxiv.org/abs/2512.14397LinksNils Thuerey and the Physics-based Simulation grouphttps://ge.in.tum.de/about/n-thuerey/Chapters00:00 Podcast intro00:39 Introducing Prof. Nils Thuerey04:13 Conversation begins05:13 From Computational Numerics to Graphics and Visual Effects07:17 Physics-Based Deep Learning Before ChatGPT10:01 CNNs, Graphics and the Move into Engineering Applications12:37 PhiFlow and Differentiable Physics14:13 Can Neural Emulators Surpass Their Training Data?18:00 The Promise and Limits of Foundation Models for PDEs20:43 Tadpole and Synthetic Online Pre-Training24:07 From Canonical PDEs to Navier-Stokes and Industrial CFD26:35 What Do Foundation Models Actually Learn?28:36 PDE Pre-Training vs. Millions of CFD Simulations33:08 Scaling 3D Transformers and Training Infrastructure35:58 Generating and Training on Data in Real Time38:00 LES, Temporal Data and Turbulence42:15 Overfitting and Correlated Simulation Data44:27 Bringing Differentiable Solvers Back into the Loop45:31 WeatherBench, APEBench and the Value of Benchmarks47:09 SuperWing, Open Datasets and Commercial Data51:31 Open Source, Commercial Models and a Technical Oscar56:17 Academia, Startups and Industry01:00:55 What Will Change Over the Next Five Years?01:02:07 World Models and the Need for Physics01:08:19 Agents, Tool Use and Calling Physics Simulators01:11:22 Career Advice for AI and Simulation01:13:54 Closing Thoughts

RANS uncertainty, data-driven turbulence modeling and AI for Science are the focus of this conversation with Professor Paola Cinnella, Professor of Fluid Mechanics at Sorbonne University and Director of SCAI. Neil and Paola discuss high-order methods, dense gases, Bayesian uncertainty, AirfRANS, surrogate modeling, scientific publishing and education in the AI era.Full episode, corrected transcript and resources:https://neilashton.co.uk/podcasts/s4-e4-prof-paola-cinnella-on-ai-for-science-and-fluid-mechanics/TopicsFluid mechanics, CFD and high-order schemesDense gases, real-gas effects and expansion shockwavesUncertainty quantification and Bayesian methodsRANS turbulence-model uncertaintyAirfRANS and CFD datasets for machine learningTurbulence modeling vs. surrogate modelingScientific publishing and ML-for-CFD standardsSCAI and AI for ScienceEducation, ChatGPT and centaur scientistsPapersQuantification of model uncertainty in RANS simulations: A review — Heng Xiao, Paola Cinnellahttps://doi.org/10.1016/j.paerosci.2018.10.001Discovery of Algebraic Reynolds-Stress Models Using Sparse Symbolic Regression — Martin Schmelzer, Richard P. Dwight, Paola Cinnellahttps://doi.org/10.1007/s10494-019-00089-xBayesian estimates of parameter variability in the k-epsilon turbulence model — W.N. Edeling, P. Cinnella, R.P. Dwight, H. Bijlhttps://doi.org/10.1016/j.jcp.2013.10.027AirfRANS: High Fidelity Computational Fluid Dynamics Dataset for Approximating Reynolds-Averaged Navier-Stokes Solutionshttps://arxiv.org/abs/2212.07564Data-driven turbulence modeling — Paola Cinnellahttps://arxiv.org/abs/2404.09074Direct numerical simulations of supersonic turbulent channel flows of dense gases — Luca Sciacovelli, Paola Cinnella, Xavier Gloerfelthttps://doi.org/10.1017/jfm.2017.237LinksPaola Cinnella named Director of SCAIhttps://scai.sorbonne-universite.fr/news/paola-cinnella-new-directorSCAIhttps://scai.sorbonne-universite.fr/Paola Cinnella — HAL publicationshttps://cv.hal.science/paola-cinnellaPaola Cinnella — Google Scholarhttps://scholar.google.com/citations?hl=fr&user=wBRA0JAAAAAJERCOFTAC SIG 54 — Machine Learning for Fluid Dynamicshttps://www.ercoftac.org/special_interest_groups/54-machine-learning-for-fluid-dynamics/master-of-science-internships/Chapters00:00 Podcast intro00:39 Introducing Prof. Paola Cinnella03:28 Conversation begins03:56 How Paola Found Fluid Mechanics07:09 Moving from Italy to France08:37 High-Order Schemes and Compressible Flows09:30 Building an Academic Career12:06 Dense Gases and Uncertainty Quantification15:16 Expansion Shockwaves and Real-Gas Effects19:17 Returning to Paris and Academic Mobility24:52 Academia, Passion and Persistence27:51 Bayesian Methods and Turbulence Uncertainty30:47 Learning Statistics Across Disciplines33:07 LearnFluidS, AirfRANS and CFD Datasets36:33 Skepticism and Physics in ML Turbulence Modeling40:41 Could ML Lead to a Universal Turbulence Model?42:59 Turbulence Models, Surrogate Models and RANS45:03 Why LES Alone Cannot Solve Optimization47:15 Multi-Fidelity Modeling49:08 What Computers & Fluids Looks for in ML-for-CFD Papers54:05 CFD Metrics vs. Machine-Learning Metrics57:13 Overselling, Publication Pressure and Quality01:02:22 SCAI and AI for Science01:06:07 Cross-Disciplinary AI for Science01:09:26 Education in the AI Era01:12:44 Critical Thinking and AI Outputs01:18:15 AI as a Companion, Not a Replacement01:21:42 AlphaFold and the Future of Discovery01:23:43 Training Centaur Scientists01:25:11 Closing Thoughts

Foundation models, explainable AI and autonomous discovery in fluid mechanics are the focus of this conversation with Professor Ricardo Vinuesa, Associate Chair for Research and Associate Professor of Aerospace Engineering at the University of Michigan. Neil and Ricardo discuss latent representations, turbulence, reduced-order modeling, flow control and whether AI can discover physical mechanisms that humans might miss.Full episode, corrected transcript and resources:https://neilashton.co.uk/podcasts/s4-e3-prof-ricardo-vinuesa-on-ai-for-fluid-mechanics/TopicsCan fluid mechanics have a “ChatGPT moment”?Foundation models and latent representations for turbulent flowsExplainable AI, causality and identifying the mechanisms that matterWhy classical coherent structures may tell only part of the turbulence storyPhysics-informed vs purely data-driven machine learningReduced-order modeling, autoencoders, transformers and nonlinear compressionDeep reinforcement learning for flow control and optimizationAgentic AI and autonomous scientific discovery in PDE-governed systemsHow academia, computer science and engineering education must adapt to AIPapersAgentic Exploration of PDE Spaces using Latent Foundation Models for Parameterized Simulations — Abhijeet Vishwasrao et al.https://arxiv.org/abs/2604.09584Multi-agent LLMs and latent foundation models autonomously explore flow physics in a tandem-cylinder problem.Enhancing computational fluid dynamics with machine learning — Ricardo Vinuesa, Steven L. Bruntonhttps://doi.org/10.1038/s43588-022-00264-7A roadmap for useful ML in CFD, including faster simulations, turbulence models and reduced-order models.Identifying regions of importance in wall-bounded turbulence through explainable deep learning — Andrés Cremades et al.https://doi.org/10.1038/s41467-024-47954-6Explainable AI identifies flow structures that matter for prediction and control.β-Variational autoencoders and transformers for reduced-order modelling of fluid flows — Alberto Solera-Rico et al.https://doi.org/10.1038/s41467-024-45578-4Disentangled latent spaces, autoencoders and transformers support interpretable reduced-order models.Improving turbulence control through explainable deep learning — Miguel Beneitez et al.https://arxiv.org/abs/2504.02354Explainable AI and deep reinforcement learning target turbulence-sustaining mechanisms.LinksVinuesaLabhttps://www.vinuesalab.com/Ricardo Vinuesa — University of Michigan Aerospace Engineeringhttps://aero.engin.umich.edu/people/ricardo-vinuesa/AI and ML for Fluid Dynamics course — Ricardo Vinuesa and Sergio Hoyashttps://www.flowthermolab.com/courses/ai-ml-for-fluids/VinuesaLab YouTube channelhttps://www.youtube.com/@VinuesaLabAI for Fluid Mechanics, Sustainability & XAI — Ricardo Vinuesahttps://www.youtube.com/watch?v=TOfwf4ffPnUModelling and controlling turbulent flows through deep learning — Ricardo Vinuesahttps://www.youtube.com/watch?v=0AOY_agZ8WMChapters00:00 Podcast intro03:20 The Evolution of Foundation Models in Fluid Dynamics10:22 Understanding Explainable AI in Fluid Mechanics15:34 Challenges in Data Fidelity for Foundation Models20:29 Machine Learning vs. Reduced-Order Modeling24:22 The Shift from Turbulence Modeling to Surrogate Models29:48 Exploring Agentic Systems for Scientific Discovery37:21 Exploring Latent Representations in Fluid Dynamics40:40 The Role of AI in Autonomous Discovery41:57 Bridging Fluid Mechanics and Computer Science45:28 Data-Driven vs. Physics-Driven Models51:34 The Role of Academia in AI and Fluid Mechanics56:27 Optimization and Control in Machine Learning01:00:28 The Future of AI in Fluid Dynamics: Beyond ChatGPT

Physics-informed AI, DMD, SINDy and data-driven engineering are the focus of this conversation with Professor J. Nathan Kutz, Director of Physics-Informed AI at Autodesk. Neil and Nathan trace machine learning’s evolution in engineering, the role of physics in trustworthy models, and the future of autonomous agents, design automation and human expertise.Full episode, corrected transcript and resources:https://neilashton.co.uk/podcasts/s4-e2-prof-nathan-kutz-on-physics-informed-ai-and-data-driven-modeling/TopicsHistory of machine learning in engineeringDynamic Mode Decomposition (DMD) and Sparse Identification of Nonlinear Dynamics (SINDy)Physics-informed AI and reduced-order modelingThe debate between physics-based and data-driven modelsThe future of autonomous agents and their impact on industryPapersFlower discrimination by pollinators in a dynamic chemical environment — Jeffrey A. Riffell, Eli Shlizerman, Elischa Sanders, Leif Abrell, Billie Medina, Armin J. Hinterwirth, J. Nathan Kutzhttps://doi.org/10.1126/science.1251041Nathan’s early move into neuroscience and data-driven biological modeling.Data assimilation and discrepancy modeling with shallow recurrent decoders — Yuxuan Bao, J. Nathan Kutzhttps://arxiv.org/abs/2512.01170Using ML to close the gap between simulation and reality.Discovering governing equations from data by sparse identification of nonlinear dynamical systems — Steven L. Brunton, Joshua L. Proctor, J. Nathan Kutzhttps://doi.org/10.1073/pnas.1517384113The foundational paper introducing SINDy.On Dynamic Mode Decomposition: Theory and Applications — Jonathan H. Tu, Clarence W. Rowley, Dirk M. Luchtenburg, Steven L. Brunton, J. Nathan Kutzhttps://doi.org/10.3934/jcd.2014.1.391A key reference for Dynamic Mode Decomposition.Data-driven discovery of partial differential equations — Samuel H. Rudy, Steven L. Brunton, Joshua L. Proctor, J. Nathan Kutzhttps://doi.org/10.1126/sciadv.1602614Extends equation discovery to PDEs and physical systems.Deep learning for universal linear embeddings of nonlinear dynamics — Bethany Lusch, J. Nathan Kutz, Steven L. Bruntonhttps://doi.org/10.1038/s41467-018-07210-0Connects deep learning with Koopman theory.Articraft: An Agentic System for Scalable Articulated 3D Asset Generation — Matt Zhou, Ruining Li, Xiaoyang Lyu, Zhaomou Song, Zhening Huang, Chuanxia Zheng, Christian Rupprecht, Andrea Vedaldi, Shangzhe Wuhttps://arxiv.org/abs/2605.15187A practical example of agentic AI for engineering design.LinksArticraft project pagehttps://articraft3d.github.io/Chapters00:00 Podcast intro00:40 Introduction to Episode05:00 Welcoming Prof. Kutz10:34 The Evolution of Data-Driven Modeling16:13 Understanding the SINDy Algorithm and Its Implications22:14 Comparing Reduced-Order Modeling and Modern Machine Learning28:29 The Role of Data in Machine Learning and Physics34:23 Challenges in Extrapolation and Real-World Applications40:46 Insights from McLaren and Team Dynamics46:07 The Shift from Academia to Industry48:53 Collaboration and Innovation in Engineering51:57 The Role of Human Expertise in Design54:45 Leveraging AI in Formula One57:32 The Future of AI and Workforce Dynamics59:06 Navigating Career Choices in a Changing Landscape01:03:02 The Evolution of Thought in Engineering01:09:06 Preparing for the Future of Technology01:14:04 Responsible Use of AI in Engineering

In this episode, Neil explores how agents, foundation models, and AI are set to transform the Computer-Aided Engineering (CAE) and Electronic Design Automation (EDA) landscapes. He shares a comprehensive historical perspective and predicts a near-future where AI-driven automation redefines engineering workflows, productivity, and innovation.Main Topics:The evolution of simulation codes from the 1960s to modern commercial softwareThe rise of cloud computing, GPUs, and their impact on CAE and EDA industriesThe integration of AI, surrogate modeling, and foundation models into simulation workflowsThe emergence of agentic AI systems capable of autonomously performing complex engineering tasksThe strategic responses of major software companies to AI and agent technologiesThe potential democratization and automation of engineering design through AI agentsCritical questions on model ownership, transparency, and industry adoptionTimestamps:00:00 - Podcast intro00:40 - Introduction: How agents and foundation models will disrupt CAE & EDA01:40 - Historical overview: From code writing in the 60s to commercial software03:10 - Growth of aerospace and automotive industry codes and commercialization04:40 - The impact of HPC, cloud computing, and hardware evolution06:25 - Rise of cloud SaaS models and "sassification" of simulation tools07:40 - Big tech entrance: AWS, Microsoft, and Google in CAE & EDA09:00 - GPU acceleration: Changed landscape in past three to four years09:10 - The role of AI startups offering surrogate models and real-time simulation10:40 - Industry consolidation: Mergers and acquisitions among software giants11:40 - The emergence of foundation models and surrogate systems in simulation13:00 - The significance of agents: Combining AI, models, and automation14:10 - Capabilities of autonomous AI agents in complex engineering workflows15:25 - Practical use cases: Running simulations, setting up experiments, and data analysis16:10 - Questions about model ownership, open-source codes, and licensing16:40 - How agent-driven automation could democratize engineering expertise19:40 - The future of AI in engineering: Collaboration, transparency, and scientific rigor21:25 - Final thoughts: Opportunities, challenges, and the transformative potential of AIPlease note that this episode expresses my personal opinion and does not represent the views of NVIDIA.Full episode, corrected transcript and resources:https://neilashton.co.uk/podcasts/s4-e1-are-ai-agents-and-foundation-models-about-to-rewrite-cae/

In this conversation, Neil Ashton and Prof. Siddhartha Mishra, and Prof. Johannes Brandstetter discuss their recent paper on AI foundation models in computational fluid dynamics (CFD). They explore the backgrounds of the speakers, the journey to writing the paper, the role of AI in CFD, and the challenges of scaling laws and data generation. The discussion also covers model training costs, open questions, and future directions for research in this field.Fluid Intelligence: A Forward Look on AI Foundation Models in Computational Fluid Dynamics : https://arxiv.org/abs/2511.20455v1Full episode, corrected transcript and resources:https://neilashton.co.uk/podcasts/s3-e9-fluid-intelligence-with-johannes-brandstetter-and-siddhartha-mishra/

In this episode, Neil Ashton discusses various conferences and workshops in the automotive, aerospace, and machine learning fields. He highlights the importance of these events for networking, education, and staying updated with industry trends. From the SAE and AIAA events to machine learning workshops, Neil provides insights into what attendees can expect and the value of participating in these gatherings.Full episode, corrected transcript and resources:https://neilashton.co.uk/podcasts/s3-e8-the-conference-connection-hpc-cae-ml-and-engineering/