Physics-Informed ML Engine Simulation
A Hamiltonian Neural Network that learns multi-body piston-crank dynamics while conserving energy by construction, used as a fast surrogate for finite-element simulation.
-
0
Numerical energy drift
-
FEM
Accuracy target
Classical neural surrogates learn to imitate a simulation and then quietly violate its physics — energy leaks, trajectories drift, and long rollouts become fiction. This project takes the opposite approach: build the conservation law into the model’s structure so it cannot be broken.
Learning the Hamiltonian, not the trajectory
A Hamiltonian Neural Network learns the dynamics of a multi-body piston-crank subsystem by modelling its energy function rather than its state transitions. Motion is then recovered from the gradients of that learned energy.
Symplectic gradients, computed automatically
Energy conservation is enforced in the model’s latent phase space through symplectic gradients obtained by automatic differentiation. Because the update respects the geometry of Hamiltonian flow, integration error stops accumulating as drift — the simulation stays physical over long horizons.
A surrogate that plugs into real solvers
The trained network is integrated with SciPy ODE solvers, so it can stand in for a finite-element model inside an existing numerical pipeline — matching FEM accuracy at a fraction of the runtime.