| name | physics-guided-neural-networks |
| description | Physics-guided neural network design and training methods. Embed physical laws, constraints, and symmetries into neural network architecture for improved modeling of physical systems (quantum mechanics, statistical physics, fluid dynamics, materials science). Activation: physics guided neural network, 物理学指导神经网络, physics-informed neural network, PINN, physics-constrained learning, quantum neural network, physics-aware training. |
Physics-Guided Neural Networks
Combine physics knowledge with deep learning for more accurate, efficient, and physically consistent models.
Core Principles
1. Physics as Architecture Constraints
Embed physical laws directly into network design:
- Symmetry preservation: Network outputs respect physical symmetries
- Conservation laws: Energy, momentum, mass automatically conserved
- Boundary conditions: Hard constraints enforced by architecture
Example: Equivariant networks that preserve rotational symmetry in molecular dynamics.
2. Physics as Loss Functions
Add physical constraints to training objectives:
- Physics-informed loss: Penalize violations of physical laws
- Energy conservation: Minimize deviation from energy conservation
- Equation residuals: Minimize PDE/ODE residuals in predictions
loss = data_loss + physics_loss_weight * physics_residual
3. Physics-Encoded Outputs
Force outputs to satisfy physical properties:
- Unit constraints: Output units match physical units
- Positive-definite: Quantities like density are always positive
- Probability distributions: Quantum states normalized and positive
Workflow
Step 1: Identify Physical Laws
What physics governs your system?
- Conservation laws (energy, momentum)
- Symmetries (rotation, translation, gauge)
- Constitutive relations (stress-strain, etc.)
- Boundary and initial conditions
Step 2: Choose Embedding Strategy
| Strategy | When to Use | Example |
|---|
| Architecture constraint | Hard physical laws | Equivariant layers |
| Loss function penalty | Soft constraints, approximate physics | PDE residuals |
| Output encoding | Physical properties must be satisfied | Positive density |
Step 3: Implement Physics Constraints
Architecture-based (most reliable):
class E3Layer(nn.Module):
...
Loss-based (flexible):
def physics_loss(prediction, x, t):
residual = compute_pde_residual(prediction, x, t)
return torch.mean(residual**2)
total_loss = mse_loss + physics_weight * physics_loss
Output-encoded (guaranteed):
density = torch.exp(raw_output)
Application Domains
Quantum Mechanics
- Neural-network quantum states: Represent quantum wavefunctions
- Quantum entanglement: Learn entanglement patterns
- Hamiltonian learning: Infer quantum Hamiltonians from data
Key paper: arXiv:2604.03482 - "Learning high-dimensional quantum entanglement through physics-guided neural networks"
Example:
psi_normalized = psi_raw / torch.norm(psi_raw)
probability = torch.abs(psi_normalized)**2
Fluid Dynamics
- Navier-Stokes solver: Learn flow fields
- Turbulence modeling: Physics-constrained turbulence
- Boundary layer: Enforce no-slip boundary conditions
Materials Science
- Crystal energy landscapes: Predict crystal properties
- Phase transitions: Learn thermodynamic transitions
- Mechanical properties: Stress-strain relationships
Key paper: arXiv:2604.04636 - "Interpretation of Crystal Energy Landscapes with Kolmogorov-Arnold Networks"
Best Practices
- Start simple: Begin with one physics constraint, add more incrementally
- Balance weights: Tune physics_loss_weight to balance data vs physics
- Verify physics: Check that trained model actually satisfies physics
- Use symbolic: Symbolic differentiation for exact PDE residuals
- Test extremes: Physics should hold even for unusual inputs
Common Pitfalls
- Over-constraining: Too many physics terms → model can't learn from data
- Wrong physics: Using incorrect physical laws
- Numerical issues: Computing derivatives numerically can be unstable
- Training instability: Physics loss gradients may conflict with data gradients
Resources
References (load as needed)
references/quantum_states.md: Neural network quantum states methods
references/pde_solver.md: Physics-informed neural networks for PDEs
references/symmetry.md: Equivariant neural networks
Key Papers
- arXiv:2604.03482 - Physics-guided quantum entanglement
- arXiv:2604.04435 - Neural-network quantum states for few-body problems
- arXiv:2604.04636 - Kolmogorov-Arnold Networks for crystal energy
- Raissi et al. (2019) - Physics-informed neural networks (PINN)
External
Activation Triggers
Use this skill when:
- Modeling physical systems (quantum, fluid, materials, etc.)
- User mentions "physics-informed", "PINN", "physics-guided"
- Data has physical constraints that should be preserved
- Need interpretable, physically consistent predictions
- Working with quantum states, entanglement, or Hamiltonians
Example Usage
User: "How do I build a neural network that respects energy conservation?"
Agent:
- Identify energy conservation as key physics
- Recommend architecture-based or loss-based approach
- Show code for energy-preserving layer or loss function
- Explain trade-offs between hard vs soft constraints
- Reference quantum states or fluid dynamics examples
Related Skills
- neural-network-quantum-states: Specialized for quantum systems
- pde-neural-solver: For differential equations
- equivariant-neural-networks: For symmetry-preserving networks