Nonlinear Solvers¶
Exasim solves nonlinear algebraic systems with Newton iteration. Nonlinearity can come from fluxes, sources, EOS closures, boundary conditions, artificial viscosity, auxiliary equations, or implicit time integration.
Newton Formulation¶
Let \(U\) denote the unknown vector for the active solver path:
- LDG: element state unknowns;
- HDG: trace unknowns after static condensation.
The nonlinear residual is
At Newton iteration \(k\), the correction \(\Delta U^k\) satisfies
where
The state is updated as
where \(\alpha\) may be reduced by the solver when damping is required.
Sources Of Nonlinearity¶
| Source | Example |
|---|---|
| Convective fluxes | Euler and Navier-Stokes equations. |
| Diffusive fluxes | Temperature- or state-dependent viscosity. |
| Source terms | Chemistry, relaxation, body forces. |
| EOS and material laws | Pressure, temperature, transport coefficients. |
| Boundary fluxes | Wall, inflow, outflow, and Riemann boundary states. |
| DIRK stages | Implicit stage residuals for time-dependent simulations. |
LDG Versus HDG Newton Systems¶
| Solver path | Newton unknown | Jacobian action |
|---|---|---|
| LDG | Element state vector | Matrix-free residual-based matrix-vector products. |
| HDG | Trace vector after condensation | Matrix-based trace operator assembled from element Jacobians. |
The nonlinear convergence criteria are controlled by fields such as
NewtonIter, NewtonTol, NLiter, and NLtol, depending on frontend or text
input path.
Convergence Diagnostics¶
Solver output typically reports the solution norm, residual norm, GMRES iterations, and Newton update quality. Use these together:
- residual norm decreasing but GMRES stagnating suggests a linear solver or preconditioner issue;
- GMRES convergence with poor Newton progress suggests nonlinear stiffness, bad initial conditions, or inconsistent model Jacobians;
- sudden residual growth often indicates invalid physics states, boundary condition errors, or excessive time step size.
Practical Guidance¶
- Start with conservative nonlinear tolerances when validating a new model.
- Verify boundary conditions and flux Jacobians before tuning GMRES.
- For stiff parameter sweeps, use warm starts only when neighboring parameter points are physically close.
- For time-dependent runs, distinguish a failed DIRK stage from a failed steady-state nonlinear solve; the remedies may differ.