DJ4Earth: Differentiable programming in Julia for Earth system models

The project (start date 2021/08), which we have since given the shorter title “Differentiable programming in Julia for Earth system models” (DJ4Earth) has as its overall goal to leverage the power of differentiable programming to seamlessly integrate gradient-based approaches in inverse modeling and machine learning. Here “differentiable programming” (DP) refers to the ability, for a given computer program representing a function, such as a neural network, or a physics-based numerical model, to generate another efficient computer program (code) representing the function’s or model's derivative (e.g., its gradient or Hessian). Ideally, the process of generating this derivative is achieved through automatic differentiation (AD), but often enhanced – or combined – with approaches of developing efficient solvers for the derivative function.