Skip to content

Introduce N-level autoregressive GP and random adaptive multifidelity workflow - #439

Draft
eparish1 wants to merge 12 commits into
developfrom
feature/multifidelity-ar-gp-contract
Draft

eparish1 wants to merge 12 commits into
developfrom
feature/multifidelity-ar-gp-contract

Conversation

@eparish1

@eparish1 eparish1 commented Nov 19, 2026 •

Copy link
Copy Markdown
Contributor

Scope (first implementation for #438)

  • Introduce MultiFidelityQoiModel and MultiFidelityQoiModelWithErrorEstimate protocols with fidelity-selectable QoIs and uncertainty indicators.
  • Implement an N-level jointly conditioned autoregressive GP with non-nested data and a builder using the generic FidelityData contract.
  • Add run_adaptive_multifidelity(...) with fixed per-level initial/batch allocations, adaptive_sampling_strategy='random', and optional nested pairing (default true).
  • Add FidelityLevel containing a QoiModel, unique name, and fixed positive nominal cost; enforce max_total_cost without dynamic cost estimation.
  • Add MultiFidelityKFoldValidation(num_folds=5): hold out highest-fidelity QoIs, retain all lower-fidelity data, rebuild per fold, and return max held-out relative-L2 error. Separately configurable sampling and validation strategies.
  • Reuse the existing blocking/nonblocking model evaluator and save per-round counts, errors, and total nominal cost.
  • Tests for 2- and 3-level training, paired/nonpaired sampling, K-fold validation, and budget enforcement.

Deferred

  • Uncertainty-driven/cost-aware acquisition; the initial 'adaptive' workflow adapts its trained surrogate and convergence status, not its random sample locations.
  • Runtime cost measurements, resumable checkpoints, vector QoIs, and shared POD representations.
  • Richer coupling structures and fidelity-dependent QoI maps.

Review

Draft pending CI and focused review. Fixed allocation charges all scheduled evaluations; no retry or restart is implemented.

Steady CDR physical-model integration

  • Reuse examples/models/steady_cdr.py and SteadyCdrQoiModel with 9×9 coarse and 21×21 fine grids; use the physically comparable scalar domain integral of the steady PDE state as QoI (the original wrapper returns grid-dependent vectors).
  • Add tests/romtools/workflows/test_multifidelity_cdr.py, exercising coupled coarse/fine solves, randomized nested evaluations, two adaptation rounds, K-fold history and independent fine-grid prediction.
  • Add examples/multifidelity_cdr.py (--smoke or default) that compares held-out fine-grid relative L2 errors for direct LF, HF-only GP and multifidelity GP and writes results to NPZ. Full benchmark uses 15×15/39×39 grids with jointly optimized GP hyperparameters.
  • Fixed costs are nominal (1 LF, 10 HF), not measured runtimes. No accuracy improvement is assumed without benchmark measurements.

CI currently rerunning after fixes for read-only GP variance and the budget-test fixture assertion.

@github-actions

github-actions Bot commented Nov 19, 2026 •

Copy link
Copy Markdown
Contributor
PR Preview Action v1.8.2

QR code for preview link

🚀 View preview at
https://pressio.github.io/rom-tools-and-workflows/pr-preview/pr-439/

Built to branch gh-pages at 2026-11-19 20:37 UTC.
Preview will be ready when the GitHub Pages deployment is complete.

@eparish1 eparish1 changed the title Introduce multifidelity QoI protocols and jointly conditioned autoregressive GP Introduce N-level autoregressive GP and random adaptive multifidelity workflow Nov 19, 2026

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant