pinn-serving: Derivative-Aware Serving of Physics-Informed Neural Networks

Abstract

A serving harness and benchmark suite for physics-informed neural networks (PINNs). Introduces derivative-graph baking, which differentiates a tanh MLP symbolically offline and emits the closed-form derivative recursion as a forward-only program, making derivative-valued queries compatible with standard inference backends and quantisation. Includes evidence that output-space accuracy monitoring cannot detect physics degradation at serving time, and a matched comparison against a Crank-Nicolson solver.