Surgical Hill-Kinetics Saturation: Eliminating Finite-Time Blowup and Adaptive Solver Grid Collapse in Neural ODEs

Abstract

Continuous-Time Neural Networks and Neural Ordinary Differential Equations (Neural ODEs) suffer from severe numerical stiffness, finite-time gradient blowup, and adaptive solver grid collapse (NFE -> infinity) when weight magnitudes accumulate during training. In this work, we propose Surgical Hill-Kinetics Field Saturation, a cross-domain operator transfer regularizer inspired by bio-enzymatic auto-activation kinetics. By selectively applying rational saturation to divergent vector fields while maintaining sub-linear manifolds, our method guarantees global C^infty smoothness, non-zero autograd gradient flow, and N-dimensional LaSalle sphere boundedness. Certified by Microsoft Z3 SMT Theorem Prover (PASSED_LASALLE_SPHERE), PyTorch Adjoint benchmarks on stiff multi-trajectory Van der Pol dynamics (mu = 3.0, hidden_dim = 128) demonstrate a > 55% reduction in prediction error (MSE = 0.8357 vs 1.8651) over 100 training epochs, eliminating trajectory collapse while incurring virtually zero computational solver overhead (NFE = 241.9 vs 230.9, < 5% difference).