Differentiable Fortran with LFortran and Enzyme
Summary
Differentiable Fortran with LFortran and Enzyme describes making legacy Fortran/C/C++ simulations differentiable by differentiating LLVM IR using Enzyme, enabling integration with JAX via Tesseract. The piece covers the motivation, pipeline, practical pitfalls (such as NaNs with -O3), and experiments including a 2D heat solver and a 900-variable inverse problem, demonstrating gradients that match analytic results and highlighting composability with Python-based ML tooling.