Hamiltonian Neural Network
A physics-informed machine learning model that embeds the physical priors of Hamiltonian mechanics into network architecture and loss functions. It is used to learn system dynamics while enforcing physical properties such as energy conservation.
Timeline (last 90 days)
Both the training and inference of Hamiltonian neural networks (HNNs) depend on the target system's generalized momenta, and this dependence severely limits their application in real-world scenarios.
A new study proposes a general extended architecture that, when the analytical form of the generalized momenta is unknown, simultaneously learns the system's generalized momenta, Hamiltonian structure and dynamics solely from data of the states and their derivatives.
HNNs and their variants, by explicitly encoding Hamiltonian mechanics, can guarantee physical properties such as energy conservation during modeling and provide interpretable information.
The new architecture treats the generalized momenta as a latent variable to be learned, jointly optimized with the Hamiltonian structure and the dynamics model.
The new architecture's input requires only the state variables and their derivatives (which may be non-sequential data), and its output includes the learned generalized momenta as well as the system's evolution laws.
For HNNs, nonholonomic constraints mean that both the definition of the generalized momenta and the Hamiltonian structure may require special handling.
The previous high barrier for HNNs was the momentum data, and the new architecture lowers this barrier to "states and their derivatives," the latter of which is usually easier to obtain in simulation and experiments.
All Facts (7)
Both the training and inference of Hamiltonian neural networks (HNNs) depend on the target system's generalized momenta, and this dependence severely limits their application in real-world scenarios.
50%UnverifiedA new study proposes a general extended architecture that, when the analytical form of the generalized momenta is unknown, simultaneously learns the system's generalized momenta, Hamiltonian structure and dynamics solely from data of the states and their derivatives.
50%UnverifiedHNNs and their variants, by explicitly encoding Hamiltonian mechanics, can guarantee physical properties such as energy conservation during modeling and provide interpretable information.
50%UnverifiedThe new architecture treats the generalized momenta as a latent variable to be learned, jointly optimized with the Hamiltonian structure and the dynamics model.
50%UnverifiedThe new architecture's input requires only the state variables and their derivatives (which may be non-sequential data), and its output includes the learned generalized momenta as well as the system's evolution laws.
50%UnverifiedFor HNNs, nonholonomic constraints mean that both the definition of the generalized momenta and the Hamiltonian structure may require special handling.
50%UnverifiedThe previous high barrier for HNNs was the momentum data, and the new architecture lowers this barrier to "states and their derivatives," the latter of which is usually easier to obtain in simulation and experiments.
50%