LFortran and Enzyme Enable Differentiable Fortran for Modern ML Pipelines

Overview of Differentiable Fortran

By combining LFortran with Enzyme, users can now achieve automatic differentiation with existing Fortran, C, or C++ code. This technology bridges the gap between decades of validated simulation software and modern machine learning pipelines, facilitating integration without the need for complete rewrites.

Technical Implementation

The process involves using Enzyme to apply automatic differentiation at the LLVM Intermediate Representation (IR) level, which allows for differentiation of any code that compiles to LLVM.

With LFortran, users can point their implementation at legacy solvers, such as a Fortran thermal solver, and retrieve exact gradients.

Challenges and Efforts Required

While the integration is promising, it is still experimental. Users may encounter issues such as gradients returning NaN and will need to manually compare LLVM IR diffs to resolve issues.

This labor-intensive process is crucial to ensure that the gradients match expected analytic results, thus validating the approach.

Impact on Scientific Computing

The capability to backpropagate through traditional Fortran codes opens new avenues for optimization and inverse problems in various fields, including CFD and climate modeling.

This innovation allows researchers to leverage existing trusted simulations while incorporating modern machine learning techniques without extensive rewrites.