Differentiable Fortran with LFortran and Enzyme

This article discusses the use of Enzyme, a tool for automatic differentiation, to get gradients out of legacy Fortran code. The author uses LFortran as a frontend to compile the Fortran code to LLVM IR, which can then be differentiated by Enzyme. The resulting differentiated code is used to solve a thermal forensics inverse problem, recovering a 900-element initial temperature field from 100 noisy sensor readings.