Artificial Intelligence for Physics and Engineering: Modeling and Simulation
||Artificial Intelligence for Physics and Engineering: Modeling and Simulation|
||NAISS Small Compute|
||Ralph Scheicher <email@example.com>|
||2023-10-12 – 2024-02-01|
Artificial Intelligence (AI) and Machine Learning (ML) techniques have emerged as promising strategies to tackle multiple physics and engineering challenges. The fast-paced advances have been observed across diverse applications, for instance, in the data-driven design and discovery of new Materials such as 2D-based structures for micro-devices, semiconductors, reactors, catalyzers, artificial leaves, fuel cells, and batteries. Additionally, it has served as a powerful tool to advance the solution of inverse problems involving numerical discretization of Partial Differential Equations (PDE), physics-informed modeling and forecasting dynamics of Multiphysics, and complex multiscale systems. Despite relentless progress, modeling and predicting the evolution of materials and nonlinear multiscale systems is still a tremendous challenge. This project aims at developing novel AI-based methods for the computer-aided discovery of materials, modeling, and simulation of complex physics and engineering problems using AI and ML techniques. It is an extension of Artificial Intelligence for Materials Discovery: Towards Neuromorphic Computing, Quantum Computing, and Catalysis to scale up the SSMD MBTR architecture with the Materials Project database. We will expand the scope and applications of the Sequential SSMD MBTR deep learning architecture by scaling up the training with more than 40,000 crystal structures. The results of these optimized architectures will be used to guide new experiments aiming at the discovery of new 2D Materials. A second goal is to investigate new approaches based on Graph Neural Networks by combining the Orbital Crystal Graph Convolutional Neural Networks with Graph Attention Models for predicting new Materials . Furthermore, this project initiative will explore challenges in Multiphysics and complex multiscale systems using Deep Reinforcement Learning, Graph Neural Networks, and Graph Diffusion Models. We will investigate the use of Deep Reinforcement Learning techniques for high-throughput prediction of properties and analysis of Energy-based devices (i.e., Fuel Cells and Batteries).
A second direction aims to investigate new Graph Neural Networks (GNN) extensions for learning complex physics simulations  based on Attention Models. This project aims to efficiently simulate particle systems and 3D datasets of fluids, solids, and other deformable materials (Master Thesis entitled Physics-based animation using Graph Neural Networks defended under my supervision at Linköping University). An extension will also explore the MeshGraphNets  simulations with Graph Attention Models for modeling complex systems, which are generally computationally expensive for predicting the dynamics of complex systems such as aerodynamics and structural mechanics. Further improvements can lead to more efficient engineering modeling tasks. A third direction aims to investigate the application of Physics-informed machine learning  and Diffusion Graph Neural Networks  in Multiphysics and Engineering problems, which generally are computationally expensive since it involves the numerical solution of Partial Differential Equations (PDE), i.e., Computational Fluid Dynamics (CFD)