Why do physics-informed constraints like Landau theory improve generalization better than pure neural approaches?

Your question touches on a real tension in machine learning for physics, but the provided archive doesn't directly address why physics-informed constraints like Landau theory improve generalization.

That said, I can sketch the general principle from what *is* here. When you add domain constraints—whether Landau symmetry arguments, conservation laws, or effective field theory structure—you're essentially encoding inductive bias: you're telling the network "solutions must respect these symmetries and principles." This shrinks the hypothesis space the network needs to search, which typically improves generalization on held-out data because the model isn't wasting capacity on solutions that violate physical law [AG-2024.12-1755].

The closest relevant finding: researchers building ML models for high-energy density physics found that when they constructed their architecture around *physical features* (shock profiles, material edges) rather than raw data, the models generalized better to new regimes [AG-2024.12-1755]. The constraint wasn't a mathematical penalty—it was baked into *what the network was asked to predict*.

More broadly, the archive emphasizes that pure neural approaches often fail at "tacit step reconstruction" and "global consistency constraints" in physics domains [AG-2026.04-565]; models trained on synthetic physics problems with explicit reasoning chains perform better when they're taught to respect conceptual structure [AG-2026.04-892]. But none of these papers directly compare Landau-constrained nets to unconstrained ones, or measure the generalization gap.