Learning Simulation Using Graph Networks

Artificial Intelligence (AI) is becoming more and more a part of the activities traditionally covered by the engineering analysis and simulation community. Recent advances in the application of AI, machine learning (deep learning) and predictive analytics, have brought these technologies to the fore in every area of industry.

This seminar hosted by the NAFEMS Americas Steering Committee brought together speakers from the end-user, consultancy, and academic industries to discuss where we are and how these technologies are being used to advance significantly the engineering analysis and simulation capabilities and approaches over the next 10 years.

Resource Abstract

While machine learning models promise to significantly speed up numerical simulations, they are often only accurate in a very narrow range around the training data. In practice, this requires running a large number of expensive reference simulations for each use case, negating the potential performance benefits of using learned methods.

In this talk we will present our recent work on neural graph architectures for learning physical dynamics end-to-end. We believe these methods are the path forward in learned simulation; compared to the prevailing CNN architectures, they have much stronger generalization and stability properties, and even allow to simulate larger and more complex setups than seen during training. We will explain the core ideas behind our models, and demonstrate several examples of learned 2D and 3D mesh-based (FEM) and particle (SPH, MPM) simulations.