Project Description
Physical models based on partial differential equations (PDE) offer strong generalization and interpretability, including ability to infer unobserved properties. Calibrating such models with observed data often requires propagation of derivatives through PDE solvers, constitutive relations (which describe complex materials), and data layers such as image processing. These components are often written in different languages and possess structure that enable more efficient and accurate differentiation. Enzyme is an exciting new language-agnostic automatic differentiation tool that operates on LLVM intermediate representation. This project will explore use of, and possibly contribution to, Enzyme to represent and compose structure/symmetry-exploiting algorithms found in data-driven PDE-based models. This work may use libraries such as libCEED, PETSc, and CRIKit.
Special Requirements
Familiarity with automatic differentiation or programming languages/compilers would be helpful.
Contact
- Jed Kallen-Brown (faculty)
- Leila Ghaffari (graduate student)