SciML linting - allow nn outputs not in params table

petab/v2/extensions/sciml_lint.py

@@ -27,6 +27,7 @@
     "CheckNeuralNetworkModel",
     "CheckSciMLConditionTable",
     "CheckSciMLParameterTable",
+    "get_nn_entity_petab_ids",
 ]
 
 #: Placeholder used in messages when a neural network has no ID.
@@ -67,7 +68,7 @@ def _nn_ids(problem: core.Problem) -> set[str]:
     return ids
 
 
-def _nn_entity_petab_ids(
+def get_nn_entity_petab_ids(
     problem: core.Problem,
 ) -> tuple[dict[str, str], dict[str, str], dict[str, str]]:
     """Classify NN entities referenced in the mapping table.
@@ -209,7 +210,7 @@ def run(self, problem: core.Problem) -> lint.ValidationIssue | None:
         condition_targets = {
             c.target_id for ct in problem.conditions for c in ct.changes
         }
-        nn_inputs, nn_outputs, nn_params = _nn_entity_petab_ids(problem)
+        nn_inputs, nn_outputs, nn_params = get_nn_entity_petab_ids(problem)
         array_input_ids = _array_input_ids(problem)
         array_param_layers = _array_parameter_layers(problem)
         array_param_petab_ids = {
@@ -333,7 +334,7 @@ class CheckSciMLConditionTable(lint.ValidationTask):
     def run(self, problem: core.Problem) -> lint.ValidationIssue | None:
         messages = []
 
-        nn_inputs, nn_outputs, nn_params = _nn_entity_petab_ids(problem)
+        nn_inputs, nn_outputs, nn_params = get_nn_entity_petab_ids(problem)
         array_input_ids = _array_input_ids(problem)
         array_param_layers = _array_parameter_layers(problem)
         array_param_petab_ids = {

petab/v2/lint.py

@@ -1127,6 +1127,8 @@ def append_overrides(overrides):
     parameter_ids -= condition_targets
 
     if problem.extensions.sciml is not None:
+        from .extensions.sciml_lint import get_nn_entity_petab_ids
+
         hybridization_targets = {
             hyb.target_id for hyb in problem.extensions.sciml.hybridizations
         }
@@ -1137,6 +1139,10 @@ def append_overrides(overrides):
         }
         parameter_ids -= hybridization_target_values
 
+        # NN outputs should not appear in the parameters table.
+        _, nn_outputs, _ = get_nn_entity_petab_ids(problem)
+        parameter_ids -= set(nn_outputs)
+
     return parameter_ids
 
 

tests/v2/test_sciml.py

@@ -401,6 +401,57 @@ def test_parameter_posterior_requires_bounds_or_prior():
     assert "net1_ps" in issue.message
 
 
+def _add_observable_consuming_nn_output(problem):
+    """Add an observable whose formula references an NN output directly."""
+    problem.add_mapping("net1_output2", "net1.outputs[0][1]")
+    problem.add_observable("fitness_obs", "net1_output2", noise_formula="0.05")
+    problem.add_measurement(
+        "fitness_obs", time=1, measurement=1, experiment_id="e1"
+    )
+    return problem
+
+
+def test_nn_output_in_observable_formula_not_required_parameter():
+    """NN outputs should not appear in the parameter table, and can appear in
+    observable formulas."""
+    from petab.v2.lint import get_required_parameters_for_parameter_table
+
+    problem = _add_observable_consuming_nn_output(_get_test_problem())
+
+    assert "net1_output2" not in get_required_parameters_for_parameter_table(
+        problem
+    )
+    assert problem.validate() == []
+
+
+def test_nn_output_in_noise_formula_not_required_parameter():
+    """Same for noise formulas."""
+    from petab.v2.lint import get_required_parameters_for_parameter_table
+
+    problem = _get_test_problem()
+    problem.add_mapping("net1_output2", "net1.outputs[0][1]")
+    problem.observable_tables[0]["B_obs"].noise_formula = "net1_output2"
+
+    assert "net1_output2" not in get_required_parameters_for_parameter_table(
+        problem
+    )
+    assert problem.validate() == []
+
+
+def test_genuinely_missing_output_parameter_still_reported():
+    """The NN-output carve-out does not mask real missing parameters."""
+    problem = _add_observable_consuming_nn_output(_get_test_problem())
+    # `scale` is not an NN entity and is not in the parameter table.
+    problem.observable_tables[0][
+        "fitness_obs"
+    ].formula = "scale * net1_output2"
+
+    results = problem.validate()
+    assert results.has_errors()
+    assert any("scale" in issue.message for issue in results)
+    assert not any("net1_output2" in issue.message for issue in results)
+
+
 # ---------------------------------------------------------------------------
 # Full-problem integration
 # ---------------------------------------------------------------------------