How do machine learning models ensure they preserve physics constraints like energy conservation during fast approximations?

This is a thoughtful question that touches on a genuine tension in modern computational science: neural networks are fast but fragile, while physical laws are slow but rigid. The preprints here address it from several angles.

Hard constraints vs. soft penalties

The most direct approach is to bake conservation laws into the network's *structure* rather than hoping it learns them. In the ICF inverse problem [AG-2024.12-1755], researchers don't train a single black-box network; instead they build a pipeline (radiograph → features → parameters) where each stage is trained on quantities that have clear physical meaning. The final parameters feed into a *traditional hydrodynamics code*, which guarantees energy and momentum conservation by construction. This is a hybrid approach: use ML for speed where it's safe (feature extraction), then hand off to physics-respecting solvers for the critical step.

A similar strategy appears in relativistic hydrodynamics [AG-2024.12-226], where neural networks replace slow root-finding algorithms in the conservative-to-primitive conversion—a step that must invert the equation of state while respecting thermodynamic consistency. The network is trained to high precision (errors ~10⁻⁷) on *correct* data generated by traditional methods, so it inherits their physical validity rather than learning it from scratch.

Learning in constrained spaces

When you can't hand off to a physics solver, you can constrain the *space* the network explores. In density functional theory [AG-2024.03-1817], variational autoencoders compress ground-state density profiles into a smaller representation space that is inherently "regular" and stable. Minimizing energy in this compressed space avoids the variational property violations (unphysical density oscillations) that plague unconstrained neural networks, while still achieving chemical accuracy. The constraint works because the compressed space is learned from physical data—it learns what realistic densities look like.

Soft symmetry penalties

Alternatively, you can add loss terms that *encourage* (rather than enforce) physical principles. The SEAL framework [AG-2025.11-1116] penalizes violations of symmetries like Lorentz invariance during training, letting the model learn whether a symmetry matters for the specific task. This is more flexible than hard constraints when real experiments have imperfections (finite detector granularity, energy thresholds) that subtly break ideal symmetries.

Physics-informed optimization

Even the training process itself can embed physics. The Energy Conserving Descent optimizer [AG-2025.01-1027] uses Hamiltonian dynamics—a classical physics idea—to stabilize neural network training for likelihood estimation. By borrowing structure from physics at the *optimization level*, not just the data level, it reduces sensitivity to random initialization and hyperparameter choices.

---

Why this matters: A network that violates energy conservation by 0.1% sounds acceptable until that error propagates through a long simulation and produces garbage. By anchoring ML to physics through structure, compressed spaces, or hybrid pipelines, you get both speed and reliability—you're not just hoping the network learned the rules.

The provided preprints don't detail a general, universal recipe for preserving all constraints in all fast approximations, so there's room for broader guidance from other sources on this topic.