romtools.workflows.greedy.run_greedy#

romtools.workflows.greedy.run_greedy(fom_model, rom_model_builder, parameter_space, absolute_work_dir=None, tolerance=None, testing_sample_size=10, random_seed=1, calibrated_error=True, *, initial_sample_size=2, max_iterations=100, initial_samples=None, create_rom_run_directories=True, absolute_greedy_work_directory=None)[source]#

Train a ROM using adaptive error-indicator-based sample selection.

The builder receives offline_data_dir, training_data_dirs, training_parameters of shape (N, parameter_dimension), and training_qois of shape (N, qoi_dimension), all in matching training order. Every rebuild includes all completed training FOM runs. Scalars become one-component QoI rows; array QoIs are flattened. Copies of the training arrays are passed so builders cannot modify the sampling pool or accumulated QoI history.

The returned ROM must implement compute_error_estimate in addition to the QoiModel methods. Directory-only builders must be adapted to the four-argument QoiModelBuilderWithTrainingData interface.

initial_sample_size selects the first training rows from the initial candidate pool (default two). max_iterations bounds the number of subsequent greedy FOM enrichments and defaults to 100; zero builds only the initial ROM. Without explicit initial samples, testing_sample_size must exceed initial_sample_size for enrichment so there is an untrained candidate. Calibration and convergence checks retain their existing behavior: the first convergence check follows one enrichment. Run directories are populated once, immediately before their model evaluation.

initial_samples, if supplied, is a nonempty (N, parameter_dimension) array of parameter rows to simulate before adaptation. Its row count overrides initial_sample_size; testing_sample_size then controls a separate adaptation candidate pool. Exact matches to accumulated training rows are excluded from this pool. No domain projection or reuse of existing FOM outputs is performed.

create_rom_run_directories=False skips only per-evaluation ROM run directory creation. Model methods still receive their nominal paths and must support in-memory execution. FOM, offline-data, log, and statistics directories remain available. Returns the final ROM and training history in a GreedyResult, including for initial-only and limit exits.

Parameters:
  • fom_model (QoiModel)

  • rom_model_builder (QoiModelBuilderWithTrainingData)

  • parameter_space (ParameterSpace)

  • absolute_work_dir (str)

  • tolerance (float)

  • testing_sample_size (int)

  • random_seed (int)

  • calibrated_error (bool)

  • initial_sample_size (int)

  • max_iterations (int)

  • initial_samples (ndarray | None)

  • create_rom_run_directories (bool)

  • absolute_greedy_work_directory (str)

Return type:

GreedyResult