S3 EP3 - Professor Johannes Brandstetter on AI for Computational Fluid Dynamics

In this conversation, Neil Ashton interviews Prof. Johannes Brandstetter, a physicist turned machine learning expert, about his journey from academia to industry, focusing on the application of machine learning in engineering and computational fluid dynamics (CFD). They discuss the Aurora project, the challenges of integrating machine learning with engineering, and the importance of data in training models. Johannes shares insights on the use of transformers in modeling, the significance of resolution independence, and the role of open-source practices in advancing the field. The conversation also touches on the challenges of founding a startup and the need for multidisciplinary collaboration in tackling complex engineering problems.

Links: 

Github: https://brandstetter-johannes.github.io
Emmi AI: https://www.emmi.ai
Google scholar: https://scholar.google.com/citations?user=KiRvOHcAAAAJ&hl=de

AB-UPT transform paper: https://arxiv.org/abs/2502.09692

Chapters

00:00 Introduction to Johannes Brandstetter
07:10 The Aurora Project and Key Learnings
11:15 Machine Learning in Engineering and CFD
17:19 Challenges with Mesh Graph Networks
20:16 Transformers in Physics Modeling
31:14 Tokenization in CFD with Transformers
39:58 Challenges in High-Dimensional Meshes
41:08 Inference Time and Mesh Generation
41:36 Neural Operators and CAD Geometry
45:59 Anchor Tokens and Scaling in CFD
48:40 Data Dependency and Multi-Fidelity Models
50:32 The Role of Physics in Machine Learning
54:28 Temporal Modeling in Engineering Simulations
56:58 Learning from Temporal Dynamics
1:00:58 Stability in Rollout Predictions
1:03:48 Multidisciplinary Approaches in Engineering
1:05:18 The Startup Journey and Lessons Learned