HamiltonianNeuralNetworks

HamiltonianNeuralNetworks learns Hamiltonian vector fields from data. Its network approximates the Hamiltonian, and the vector field is the symplectic gradient of the network [1].

The package provides

the architecture StandardHamiltonianArchitecture,
the losses HNNLoss, for vector field data, and SymplecticEulerLoss, for pairs of states one timestep apart,
the SymbolicPullback of HNNLoss, a gradient that SymbolicNeuralNetworks generates in closed form.

The networks train with the DataLoader, Batch, train! and the optimisers of GeometricMachineLearning; the tutorial shows the whole workflow.

HamiltonianNeuralNetworks — Module

HamiltonianNeuralNetworks

Hamiltonian neural networks: an architecture whose network is a Hamiltonian and whose vector field is its symplectic gradient, the losses that train it, and its symbolic pullback.

source