Autodiff: Why Earth System Modeling Still Depends on Compiler Infrastructure
Published:
Large scientific simulation codes written in Fortran underpin much of climate science, weather prediction, and geophysical modeling. Increasingly, these codes are not only used for forward simulation but also for inverse problems where one wants to infer parameters, sensitivities, or initial conditions from observations. Recent work in Earth system modeling has increasingly argued that differentiable programming could enable more systematic calibration, data assimilation, uncertainty quantification, and hybrid ML–physics workflows in large scientific models. In this context, automatic differentiation (AD) has become a key enabling technology. However, applying these ideas to large Fortran codebases exposes a less visible dependency: the maturity of the compiler infrastructure itself. This article discusses autodifferentiation in the context of Fortran codebases as well as its relationship to compiler infrastructure.