Show HN: Physics written into the architecture (learned Hamiltonian)
Category: ai-ml
Tags: physics-informed-ml, hamiltonian-neural-networks, liquid-neural-networks, symplectic-integrators, time-series
Score: 6.3/10 (Innovation: 7, Technical: 7, Documentation: 6, Utility: 5)
AwareLiquid-Physic combines a liquid time-constant (LTC) recurrence with a learned Hamiltonian and symplectic integrator to enforce energy conservation by architectural design, rather than as a soft loss term. The liquid core infers per-trajectory hidden parameters, conditioning the Hamiltonian's potential to adapt to families of physical systems. Benchmarks show improved long-horizon stability over unstructured baselines and static Hamiltonian models.
Target audience: machine learning researchers, computational physicists
Repository: https://github.com/AwareLiquid/AwareLiquid-Physic · Python · 2 stars
View on Hacker News