This workflow uses high-fidelity CFD or FEA to generate a design space, then trains a surrogate to make rapid estimates inside a documented validity envelope. The surrogate accelerates screening; it does not replace engineering sign-off.
Physics baseline
Choose reference cases with measured or accepted outputs. Validate mesh, material/boundary conditions, solver tolerances and convergence on AxiKeystone. Record numerical results before parallel optimisation so faster runs cannot hide changed physics.
Scale-out decision
Test domain decompositions on AxiVector and measure compute, MPI and I/O time. Use AxiLattice when design-of-experiments throughput and several users justify a scheduler. Many independent one-node cases may outperform one poorly scaling three-node case.
Dataset design
Sample the parameter space deliberately, include boundaries and known difficult regimes, and preserve failed/non-converged cases as information. Store geometry/version, mesh, solver configuration, residuals and outputs with each row. Split validation by geometry family or time when appropriate to prevent leakage.
Surrogate deployment
Expose prediction, uncertainty and validity-domain checks together. Route out-of-domain or high-consequence candidates back to the solver. Version the model with the simulation dataset and retrain only through controlled workflows. Keep plots and reports labelled as surrogate or full-order results.
Acceptance
Measure full-order numerical agreement, multi-node efficiency, campaign throughput, surrogate error by operating regime and false confidence outside the training domain. Demonstrate restart, dataset lineage and the ability to reproduce a selected design from source parameters.
Primary technical references
- OpenFOAM: running applications in parallel
- PETSc getting started
- PETSc GPU roadmap
- Slurm MPI Users Guide
References are provided for software architecture and implementation planning. Validate the versions, licences, support matrix and regulated-use requirements applicable to the final deployment.