Physics Priors are inductive biases deliberately embedded into neural network architectures, loss functions, or training procedures to ensure that model outputs respect known physical laws — conservation of energy, conservation of momentum, rotational symmetry, translational invariance, and other fundamental constraints — guaranteeing that the AI cannot produce physically impossible predictions regardless of what data it is trained on, transforming the network from an unconstrained function approximator into a physics-compliant reasoning system.
What Are Physics Priors?
- Definition: A physics prior is any architectural design choice, loss term, or training strategy that encodes known physical knowledge into a machine learning model. The term "prior" comes from Bayesian statistics — it represents what we know about the world before seeing any data, restricting the model's hypothesis space to physically plausible solutions.
- Hard vs. Soft Constraints: Hard constraints are enforced architecturally — the network structure makes it mathematically impossible to violate the physical law (e.g., Hamiltonian Neural Networks conserve energy by construction). Soft constraints are enforced through loss penalties — the training loss includes terms that penalize physical violations, guiding the model toward compliant solutions without absolute guarantee.
- Hierarchy of Physical Knowledge: Physics priors range from fundamental (energy conservation, symmetry groups) to domain-specific (material constitutive relations, fluid boundary conditions) to empirical (scaling laws, dimensional analysis). Stronger priors provide more constraint but require more domain expertise to formulate.
Why Physics Priors Matter
- Long-Term Stability: Standard recurrent neural networks trained on dynamical systems accumulate errors over time — energy drifts, trajectories diverge from physical reality, and the simulation eventually produces nonsensical states. Physics priors (particularly energy conservation through Hamiltonian structure) prevent this drift, enabling stable long-horizon predictions that track the true physical trajectory.
- Data Efficiency: Physics priors reduce the effective dimensionality of the learning problem by eliminating unphysical solutions from the hypothesis space. A model that must conserve energy has fewer valid solutions to search through, converging faster from less data than an unconstrained model.
- Scientific Trust: Scientists and engineers will not adopt AI predictions for safety-critical applications (aircraft design, nuclear reactor simulation, drug molecule design) unless the model provably respects fundamental physical constraints. Physics priors provide this guarantee, bridging the trust gap between ML predictions and engineering decisions.
- Extrapolation: Standard neural networks are unreliable outside their training distribution. Physics priors anchor the model to laws that hold universally, providing more reliable predictions in novel regimes — a Hamiltonian network trained on low-energy pendulum swings can extrapolate to high-energy regimes because energy conservation holds everywhere.
Physics Prior Implementations
| Prior | Physical Law | Implementation |
|-------|-------------|----------------|
| Hamiltonian NN (HNN) | Energy conservation | Network learns $H(q,p)$; dynamics derived from Hamilton's equations |
| Lagrangian NN (LNN) | Principle of least action | Network learns $mathcal{L}(q,dot{q})$; Euler-Lagrange equations derive motion |
| Equivariant CNN | Rotational symmetry | Group convolution guarantees equivariance to rotation group |
| Divergence-Free Networks | Mass/volume conservation | Network output constrained to have zero divergence |
| Symplectic Integrators | Phase space volume preservation | Integration scheme preserves Hamiltonian structure |
Physics Priors are guardrails for neural computation — architectural constraints that prevent AI from hallucinating unphysical behavior, ensuring that learned models play by the same thermodynamic, mechanical, and symmetry rules as the physical universe they are modeling.