PIKS: New kernel methods match PINNs with proven theoretical guarantees
Universal consistency proven for physics-informed kernel methods, closing theory gap.
Deep Dive
PIKS (Physics-Informed Kernel methodS) is a new class of models that incorporate physical laws via differential operators. Unlike PINNs, PIKS offers closed-form solutions and analytical tractability. The paper proves universal consistency for linear constraints using universal kernels (Gaussian, Matérn) and provides finite-sample bounds under suitable source conditions. Numerical experiments show PIKS can be competitive with PINNs and traditional finite element methods.
Key Points
- PIKS provides universal consistency for linear differential constraints using Gaussian or Matérn kernels.
- Closed-form solutions and finite-sample convergence bounds make PIKS more analyzable than PINNs.
- Experiments show PIKS matches or outperforms PINNs and FEM on benchmark physics problems.
Why It Matters
Bridges theory and practice in physics-informed ML, offering reliable, provably consistent alternatives to black-box neural networks.