Physics-informed operator learning for nonlinear PDEs gains an error guarantee: the first-order least-squares residual matches the error near a solution branch
Synopsis
Physics-informed operator learning trains on a PDE residual instead of solution snapshots, but whether a small residual means a small error was unproven for nonlinear equations. Writing the state as a first-order system of flux and primal variable, this work proves that the least-squares residual is equivalent to the error in the equation's own norm near a solution branch, provided the linearized residual is stable and its derivative Lipschitz continuous, and verifies the conditions for monotone reaction, solution- or gradient-dependent diffusion, and compressible Neo-Hookean hyperelasticity. A reduced basis neural operator trained on this loss kept the ratio of error to the square root of loss excess at medians comparable to the linear case.
Interpretation
Physics-informed operator learning now gives two-sided control between residual and error for a class of nonlinear parametric equations, not just a falling training loss. Earlier equivalence results held only for linear problems; nonlinear equations may admit multiple solutions, their linearizations may become singular, and constitutive laws restrict admissible states, so a small residual did not guarantee a small error. The stability and Lipschitz conditions were checked for monotone cubic reaction, Kirchhoff-type and regularized p-Laplace diffusion, and compressible Neo-Hookean hyperelasticity; the first two hold on every bounded set of states, hyperelasticity only in a neighborhood of the branch.
The training loss is assembled directly by a finite element library, and network outputs are constrained to a conforming finite element space, so fluxes and stresses are direct outputs rather than postprocessed quantities. In the linear case the loss is a quadratic form assembled once, but a cubic reaction makes it a degree-six polynomial in the reduced basis coefficients, and the Neo-Hookean law makes it nonpolynomial, so no fixed tensor can be preassembled. At each predicted state one assembly of the loss and one of a linear form are performed, with the derivative generated by symbolic differentiation and compiled once before training, then propagated to network parameters by the chain rule without forward or adjoint solves; one loss and gradient evaluation costs between a few percent and a few tenths of a finite element solve.
Perspective
The results concern parametric nonlinear PDEs written as first-order systems, with states in a conforming pair of Raviart-Thomas flux and continuous Lagrange primal elements, and are tested on three benchmarks: heat conduction with a cubic reaction over sixteen square inclusions with contrast up to one hundred, quasilinear diffusion with a lognormal permeability, and a compressible Neo-Hookean body under traction.
Hyperelastic equivalence rests on the assumption that the branch tangent operator is stable, for which no analytic lower bound is given and discrete eigenvalues serve only as numerical support; the loss does not select a branch, and hyperelasticity has no computable error bound; the computable bound for the semilinear model combines worst-case constants and is sensitive to the conductivity lower bound and reaction strength.
