Conservation-Law-Constrained Networks

These networks are built so their predictions satisfy conservation laws exactly, rather than only approximately through a penalty.

Soft versus hard constraints

A network can be encouraged to respect a conservation law by adding a penalty for violating it, a soft constraint that is only approximately satisfied. A stronger approach builds the law into the architecture so it holds exactly by construction, a hard constraint. Hard constraints guarantee physical consistency even far from the training data, which soft penalties cannot promise.

Enforcing conservation by construction

A common technique predicts fluxes rather than the quantity itself, then computes the change in the conserved quantity as the divergence of those fluxes. Because a divergence form conserves the integral exactly in a discrete sense, the total mass, momentum, or energy is preserved no matter what the network outputs. The network is free to be wrong about details but cannot break the balance.

Symmetry and invariance

Noether's theorem ties conservation laws to symmetries: energy conservation to time-translation invariance, momentum to space translation, angular momentum to rotation. Building the corresponding symmetry into a network, so its predictions transform correctly under these operations, can enforce the associated conservation law and improve data efficiency at the same time.

Trade-offs

• Hard constraints guarantee physical validity but restrict the architecture
• Soft penalties are simple and general but only approximate
• Flux-form and Hamiltonian structures cover common conservation laws
• Some constraints are difficult to impose exactly and default to penalties

Why it matters

In long simulations, small conservation errors accumulate until the result drifts far from physical reality. A network that conserves the right quantities exactly stays trustworthy over long horizons. For energy systems, where balances of particles and energy are central, respecting conservation is not a nicety but a requirement for a model to be taken seriously.