stack_scan_sensitivities: numpy.stack raises 'all input arrays must have the same shape' on a parameter_scan, aborting the gradient fit

[wshlavacek]

A gradient (trf) fit of a job containing a type: parameter_scan experiment aborts while combining the per-dose sensitivity arrays:

File "pybnf/bngsim_model/net_model.py", line 447,  in execute
File "pybnf/bngsim_model/net_model.py", line 661,  in _execute_actions
File "pybnf/bngsim_model/net_model.py", line 1192, in _run_parameter_scan
File "pybnf/data.py",                    line 86,  in stack_scan_sensitivities
File "numpy/_core/shape_base.py",        line 458, in stack
    raise ValueError('all input arrays must have the same shape')

pybnf.printing.PybnfError: Model egfr_simpull: simulation failed while computing
forward output sensitivities for gradient-based fitting. The model may use
discrete events or otherwise non-differentiable constructs.

The failure is entirely inside PyBNF: bngsim returned the per-point tensors and stack_scan_sensitivities could not combine them. numpy.stack requires exact shape agreement, so some dose in the scan produced a sensitivity array of a different shape from the others.

Reproducer

Salazar-Cavazos-2019/egfr_simpull from the PyBNF jobs collection, with the committed config switched to the gradient path:

job_type = trf
n_starts = 10
bngl_backend = bngsim

The job has two experiments — one type: parameter_scan, t_end: 300 and one plain time course:

experiment: dose_resp, type: parameter_scan, t_end: 300, data: dose_resp.exp
experiment: egf25nm, data: EGF_25nM.exp

Fails after ~11 s, before completing a single start.

Note on the wrapper message

The PybnfError blames "discrete events or otherwise non-differentiable constructs", which sent me looking at the model first. This model has no events — the cause is a shape mismatch in PyBNF's own stacking. Worth either narrowing that message or letting a ValueError from stack_scan_sensitivities surface with its own explanation, since the current text points at the wrong layer.

Suggested direction

Whatever stack_scan_sensitivities receives, the per-dose arrays are not uniform — most likely differing timepoint counts across scan points, or a mismatch between the scan experiment's grid and the plain time-course experiment's. Reporting the offending shapes (dose index, expected vs actual) before stacking would make this diagnosable from the log rather than from a numpy traceback.

Environment

PyBNF 1.6.0, bngsim 0.11.35 (v0.11.35-77-g32395e4), Python 3.12, macOS 15.

Related: this is one of two gradient-path failures I hit on parameter_scan experiments; the other is lanl/bngsim#81 (a scored scan after a pre-equilibration is refused outright).