Overview
My FYDP-II thesis code, public and evidence-first, now defended with the manuscript submitted to ICEEICT 2027. A physics-informed neural network solves the Lorenz-1960 system from the equations alone: the loss is the ODE residual, the reference solver runs only after training, and an 11,283-parameter network reaches within 3e-05 of a high-accuracy RK4/DOP853 baseline in under six minutes on a laptop CPU. Around the run of record sit a nine-architecture sweep and a windowed causal-weighting study for longer horizons. The repo preserves an immutable run catalog with settings hashes, per-collocation-point loss exports audited by hand, a two-branch source-code audit with repair notes, and provenance docs covering sixteen paper runs.