Constraints in a Differentiable Graph Solver for Modeling and Inverse Problems

19 November 20260 views

An implicit graph solver is introduced that supports end-to-end differentiation and accounts for constraints represented by a large system of equations. It has been tested on forward transient and steady-state simulations, coefficient and right-hand-side recovery, as well as learnable graph coarsening; the results indicate lower forward-simulation costs and more efficient execution of inverse problems.

Constraints in a Differentiable Graph Solver for Modeling and Inverse Problems

The paper “Automatic Differentiation with Constraints in Forward and Inverse Problems” (arXiv:2610.27888) describes a differentiable graph solver with a constraint. Its authors are Nikolay Yavich and coauthors. Here, the constraint is a large system of equations solved during numerical optimization of partial differential equations (PDEs).

The Constraint as Part of the Problem Formulation

Numerical optimization for stiff evolutionary and steady-state PDEs often requires solving a large system of equations. The paper treats this system as a constraint. The constraint is part of the problem formulation, rather than something added to a completed solution.

  • PDEs define an evolutionary or steady-state problem.
  • An approximate solution is constructed on an unstructured mesh.
  • The mesh can be represented as a graph.
  • The large system of equations is treated as a constraint.

Practical criterion: if the optimization relies on a large system of equations, the constraint should be part of the problem formulation.

Implicit Graph Solver and Differentiation

The authors present a fully differentiable implicit graph-based solver.

The solver incorporates the constraint in the form of a large system of equations. It preserves differentiability at the implicit step.

  • It is based on a graph representation of the approximate PDE solution.
  • The implicit scheme incorporates the constraint.
  • Full differentiability allows parameters to be optimized through the implicit solution.
  • Stability supports optimization.

Takeaway: differentiability at the implicit step matters for inverse problems.

Forward Modeling

The solver’s performance is evaluated on four tasks. The first group covers forward modeling, including linear and nonlinear formulations, as well as evolutionary and steady-state problems.

  • linear forward modeling;
  • nonlinear forward modeling;
  • evolutionary problems;
  • steady-state problems.

Practical criterion: the forward modeling group is relevant when solving a PDE subject to a given constraint.

Inverse Problems

The next two groups are inverse problems. They differ in the unknown quantity.

Inverse problem typeWhat the solver estimates
Coefficient-basedequation coefficients
Right-hand-side-basedthe equation’s right-hand side

Takeaway: the solver is tested on two different types of unknowns. Practical criterion: use it to estimate PDE coefficients or the right-hand side.

Learned Graph Coarsening

The fourth group is learned graph coarsening. It accelerates forward modeling.

  • the graph is coarsened using a learned method;
  • the goal is to accelerate forward modeling;
  • coarsening is included in the solver evaluation.

Practical criterion: coarsening is used when a forward run needs to be accelerated in a graph-based formulation.

Scope and Selection Criteria

The paper focuses on applications in physics and engineering. It specifically mentions geophysics, where inverse problems are solved to reduce model uncertainty.

Stability and differentiability reduce the complexity of forward modeling. They also help solve related inverse problems efficiently.

  • PDEs on an unstructured mesh;
  • graph representation of the solution;
  • a large system of equations as a constraint;
  • forward and inverse problems;
  • learned graph coarsening.

Practical criterion: this approach is suitable for PDEs on an unstructured mesh with a large system of equations as a constraint. For problems outside this framework, the paper provides no basis for choosing it.

Frequently asked questions

Constraints in a Differentiable Graph Solver for Modeling and Inverse Problems