Variationally correct physics-informed operator learning for nonlinear partial differential equations
Organizations: School of Computational Science and Engineering, Georgia Institute of Technology, Atlanta, GA 30332, USA · Department of Mathematics, University of South Carolina, Columbia, SC 29208, USA
Abstract
Physics-informed operator learning trains neural operators on the residual of a parametric partial differential equation (PDE) instead of labeled solution data, but its reliability depends on a variationally correct residual: one that is equivalent to the error in a norm dictated by the equation, the PDE-compliant norm. For nonlinear equations written as first-order systems, we prove that the first-order system least-squares residual is equivalent to the error in this norm near a solution branch if the linearized residual is stable and the residual derivative is Lipschitz continuous, possibly in a norm stronger than the error norm. We verify these conditions for monotone reaction, solution- and gradient-dependent diffusion, and compressible Neo-Hookean hyperelasticity. The stability mechanisms vary across these classes, as do the regions where the equivalence holds, from arbitrary bounded sets to a neighborhood of the branch. In a conforming finite element space, we show that the loss excess of a state near the finite element solution is equivalent to its squared distance from that solution. We use this loss to train a reduced basis neural operator with values in this space, with solution snapshots needed only for the basis and the loss and its symbolically derived gradient assembled by the finite element library. At inference, this loss estimates the error. Numerically, the square root of the loss excess tracks the error within factors comparable to those of the linear case, and the operator achieves higher accuracy in the PDE-compliant norm than grid-based baselines evaluated in the same finite element space.
Figures & tables
| Notation | Meaning |
| , | State space with the PDE-compliant norm ( 2 ); residual space |
| , | Lifted state, with flux or stress and primal variable ; selected solution branch |
| , ; , | Dirichlet and flux lifts; flux-free splitting ( 3 ) of the source |
| , ; , , | Residual and loss ( 5 ); linear part, nonlinear part, and data of the residual |
| Derivative of the residual at a state with primal component | |
| , , | Regularity space; closed -ball of radius about the branch |
| Semilinear diffusion–reaction ( Section 4.1 ) | Quasilinear diffusion ( Section 4.2 ) | Compressible hyperelasticity ( Section 4.3 ) | |
| Nonlinear term | reaction | flux | stress |
| It enters | the lower-order term | the principal part | the principal part; defined only for |
| Norm of the neighborhood ( Assumption 2 ) | , the norm of the error | , with | |
| Stability of the linearization ( Assumption 1 ) | proved: monotonicity ( Proposition 4.1 ) | proved: factorization of the tangent ( Proposition 4.4 ) | assumed: coercive or invertible tangent at the branch ( Proposition 4.5 ) |
| Estimate holds on | every bounded set in | every bounded set in | a neighborhood of the branch |
| Computable error bound | yes ( Corollary 4.2 ) | given a bound on the solution in | no; and are not known |
| Semilinear diffusion–reaction | Quasilinear diffusion | Compressible hyperelasticity | |
| Problem | |||
| Nonlinearity | reaction term , | diffusion coefficient | stress law (Neo-Hookean) |
| Equation | |||
| Domain | |||
| Linear counterpart [ 14 ] | heat conduction | Darcy flow | linear elasticity |
| Parameter | |||
| Benchmark | Training (h) | Loss and gradient (s) | Prediction (s) | FE solve (s) |
| Semilinear diffusion–reaction, | ||||
| Quasilinear diffusion | ||||
| Compressible hyperelasticity |
| Benchmark | median | range | median | range | Floor share |
| Semilinear diffusion–reaction, | – | – | |||
| Semilinear diffusion–reaction, | – | – | |||
| Quasilinear diffusion | – | – | |||
| Compressible hyperelasticity | – | – | |||
| Benchmark | States | (all states) | (at ) | (at ) |
| Semilinear diffusion–reaction, | ||||
| Quasilinear diffusion | ||||
| Compressible hyperelasticity |
| Method | Relative error | Relative error | Residual loss | |||
| Semilinear diffusion–reaction, | ||||||
| PCA-Net [CG CG 1 ] | ( | ) | ( | ) | ( | ) |
| PCA-Net [RT CG 2 ] | ( | ) | ( | ) | ( | ) |
| FNO [CG CG 1 ] | 6.26e-03 ( | 7.05e-04) | ( | ) | ( | ) |
| FNO [RT CG 2 ] | ( | ) | ( | ) | ( | ) |
| FC-PINO [CG CG 1 ] † | ( | ) | ( | ) | ( | ) |
Appendix figures & tables11 assets
Supplementary material from the paper’s appendix.
Appendix
| Semilinear diffusion–reaction | Quasilinear diffusion | Compressible hyperelasticity | |
| input/output dimension | |||
| hidden dimensions | |||
| activation function | SELU | ||
| PCA coefficients | standardized | not standardized | |
| numerical precision | float32 (PCA bases computed in float64) | ||
| number of trainable parameters | M | M | M |
| Semilinear diffusion–reaction | Quasilinear diffusion | Compressible hyperelasticity | |
| number of modes | |||
| in/out channels ( , , ) | |||
| hidden channels | |||
| FNO layers | |||
| activation | GELU | ||
| numerical precision | float32 | ||
| Semilinear diffusion–reaction | Quasilinear diffusion | Compressible hyperelasticity | |
| Architecture | |||
| number of modes | |||
| in/out channels ( , , ) | |||
| hidden channels | |||
| FNO layers | |||
| FNO activation | GELU | ||
Physics-Informed Neural Operators
Physics-informed neural operators learn mappings between physical fields while embedding governing equations or structural constraints for faster, more reliable simulation.
See all 77 papers