Differentiable Simulation
Compiling a Physics Engine into One Kernel: jit, grad and vmap in Practice
A case study of jax-spring-sim: forces derived by autodiff from a single energy function, a rollout compiled into one XLA kernel, and measured numbers for what compilation and reverse-mode differentiation actually cost.
Solving Inverse Problems in Engineering via Hardware-Accelerated Differentiable Simulations
A case study of the Differentiable Digital Twin: minimising the compliance of a loaded structure by backpropagating through a finite-element solver, with measured solve counts against finite differences, CMA-ES and random search. Updated with a matrix-free sparse solver that holds memory flat where the dense matrix alone would need gigabytes.