Building Foundation Models for Engineering Physics
We are developing scalable Physics AI architectures that learn from diverse simulation data and generalize across engineering geometries and physical regimes. That means asking bold questions and taking on hard problems: how one model can predict the physics of shapes and meshes it has never seen, and how its predictions stay faithful to the governing equations.
Why this problem is hard
Applying AI to a new engineering problem today usually means running many simulations, assembling a dataset, and training a model for that problem alone. Every new geometry starts the cycle again.
The generalization challenge
Take geometry A with roughly a thousand mesh points and geometry B with ten thousand. Different shape, different mesh, different topology. The research question is whether one model architecture can take either and return a physically meaningful prediction, without being retrained for each.
That is an open problem, and we are treating it as one.
What we are investigating
Everything we research builds toward physics foundation models: one model, pretrained on diverse simulation data, that adapts to new geometries and physics instead of being trained from scratch for every problem.
Operator learning
Neural operators learn the mapping from inputs such as geometry and boundary conditions to the full solution field. Fourier Neural Operators (FNO) do this in frequency space, making them fast and able to predict at resolutions they were not trained on.
Geometric deep learning
Graph neural networks treat a mesh as a graph and learn the physical interactions between neighboring nodes. One model can then work on any mesh, whatever its size or topology.
Transformer-based physics learning
Attention lets every point in a domain exchange information with every other, capturing long-range physical effects. We are exploring transformers as a scalable backbone for physics foundation models.
Generalization across geometries and meshes
One geometry may have a thousand mesh points, another ten thousand, with entirely different topology. Making a single model predict accurately across both is the central problem we are solving.
Physics-aware machine learning
A model that fits the data but breaks conservation laws fails in ways that look plausible. We build physical structure into our models so their predictions respect the governing equations.
Research areas
Where physics-aware modelling still has genuinely unsolved problems, rather than where it is already engineering.