Differentiable Fortran with LFortran and Enzyme

Researchers have developed a method to automatically generate gradients from legacy Fortran, C, and C++ simulation code by using Enzyme (an automatic differentiation tool) at the LLVM compiler level, allowing decades-old physics code to be integrated into modern machine learning frameworks like JAX and PyTorch without rewriting. The approach avoids the traditional options of hand-writing adjoints, using slow finite differences, or completely rewriting code, enabling new possibilities like solving inverse problems that weren't previously feasible.