Add model-selection strategies and local BioModels batch mode

README.md

@@ -40,6 +40,26 @@ uv run python data2sbml.py \
   --maxsize 24
 ```
 
+### Model-selection strategies
+
+The final inferred model can now be selected with three strategies ordered from
+cheapest to most expensive:
+
+- `--selection-strategy pysr_model_selection`: choose one safe equation per species using PySR's local `model_selection` rule.
+- `--selection-strategy global_rmse`: shortlist per-species candidates, reconstruct full models, and choose the model with the lowest simulated mean RMSE.
+- `--selection-strategy global_multiobjective`: search full reconstructed models and choose the model with the lowest `rmse_mean + penalty * total_complexity`.
+
+Example:
+
+```bash
+uv run python data2sbml.py \
+  --sbml demo_polynomial_2d.xml \
+  --species x y \
+  --selection-strategy global_rmse \
+  --selection-top-k 3 \
+  --selection-max-model-evals 64
+```
+
 ### Harder example
 
 The BioModels file `BIOMD0000000346_url.xml` is also runnable with the same pipeline, but it is a harder symbolic-regression problem and should be treated as a stress test rather than the baseline validation case.
@@ -52,11 +72,34 @@ To fetch curated BioModels identifiers via `biocompose`, cache their SBML/SED-ML
 uv run python biomodels_batch.py --limit 25 --max-species 6
 ```
 
+To run the batch directly over a local directory of downloaded BioModels files such as `artifacts/all_curated_biomodels`:
+
+```bash
+uv run python biomodels_batch.py \
+  --local-models-dir artifacts/all_curated_biomodels \
+  --limit 25 \
+  --max-species 6
+```
+
+When a model has no `.sedml` file, or the SED-ML does not expose a usable `UniformTimeCourse`, the batch now falls back to configurable simulation settings instead of skipping the model:
+
+```bash
+uv run python biomodels_batch.py \
+  --local-models-dir artifacts/all_curated_biomodels \
+  --fallback-start 0 \
+  --fallback-duration 20 \
+  --fallback-points 801
+```
+
+Use `--require-sedml` if you want to keep the previous strict behavior and fail on models without a usable SED-ML file.
+
 The batch summary is written to `artifacts/biomodels_batch_summary.tsv` and now includes:
 
 - `original_model_equation_count`
 - `inferred_model_pysr_complexity`
 - `inferred_model_node_count`
+- `simulation_settings_source`
+- `simulation_settings_note`
 
 To only cache all curated BioModels files without running symbolic regression: