An asymptotic-preserving adjoint unified gas kinetic scheme for sensitivity analysis

What the math gives to ML

The paper's transferable asset is an asymptotic-preserving discrete-adjoint construction for systems with fast microscopic relaxation coupled to slow macroscopic variables. Its key idea is to project sensitivities onto conserved macroscopic moments, solve the stiff-free macroscopic adjoint, and reconstruct the microscopic component only where needed, avoiding a backward-pass step-size restriction proportional to the relaxation time. A neural analogue is a stiff relaxation block or implicit neural layer with slow state W and fast state f, equipped with a projected custom backward pass rather than naive backpropagation through many tiny relaxation steps. The paper's residual recomputation strategy also suggests reducing activation memory in long differentiable dynamical models.