Foundation models for unstructured physics simulations

This project develops a scalable foundation model for scientific machine learning on unstructured data, focusing on point cloud representations of physical systems. Building on the Tadpole framework, it integrates transformer-based architectures with online PDE data generation to enable training on effectively unlimited datasets. The approach targets complex geometries encountered in aerodynamics and structural simulations, overcoming limitations of structured grid methods. By combining physics-based modelling and large-scale HPC training, the project aims to deliver transferable, data-efficient models. The resulting framework will accelerate simulation-driven design and establish a general paradigm for AI-driven scientific computing across engineering domains.