openmdao-sim

name	openmdao-sim
description	Use when driving OpenMDAO (NASA's open-source multidisciplinary design, analysis, and optimization framework) via Python scripts — coupled systems with MDA, gradient-based / gradient-free optimization, derivative computation, surrogate models — through sim runtime one-shot execution.

openmdao-sim

You are connected to OpenMDAO via sim-cli.

OpenMDAO is the canonical open-source MDAO framework, originally from NASA Glenn. Pure Python (numpy, scipy, networkx). Pip-installable (pip install openmdao).

Scripts are plain .py:

import openmdao.api as om

prob = om.Problem()
prob.model.add_subsystem('comp', MyComp(), promotes=['*'])
prob.driver = om.ScipyOptimizeDriver()
prob.model.add_design_var('x', lower=-10, upper=10)
prob.model.add_objective('y')
prob.setup()
prob.run_driver()
print(prob['x'], prob['y'])

Same subprocess driver mode as PyBaMM / SfePy.

---

base/ — always relevant

Path	What's there
base/reference/workflow.md	Problem → model → components → solvers/drivers → setup → run.
base/reference/components.md	ExplicitComponent, ImplicitComponent, ExecComp, IndepVarComp.
base/reference/solvers.md	NonlinearBlockGS, NewtonSolver, DirectSolver, LinearBlockGS.
base/reference/optimization.md	ScipyOptimizeDriver, design vars / objective / constraints.
base/snippets/01_sellar.py	Verified Sellar coupled MDA E2E.
base/known_issues.md	Cyclic groups need NL solver, set_input_defaults vs setup.

sdk/3/ — OpenMDAO 3.x

• sdk/3/notes.md — version-specific surface notes.

---

Hard constraints

1. Cyclic groups MUST set a nonlinear solver. Sellar-style coupled disciplines (D1 → D2 → D1) need group.nonlinear_solver = om.NonlinearBlockGS() (or NewtonSolver). Without it the residual loop is open and run_model() errors.
2. setup() is mandatory before run_model() or run_driver(). It builds the data passing graph and allocates arrays.
3. Use set_input_defaults for promoted inputs, set after add_subsystem and before setup. Setting inputs after setup uses prob['x'] = ....
4. Acceptance != "ran without error". Always validate against a known reference (Sellar has y1=25.588 / y2=12.058 by convention; gradient-based runs should report objective improvement).
5. Print results as JSON on the last stdout line.

---

Required protocol

1. Gather inputs:
   • Category A: disciplines / equations, design variables, objectives, constraints, acceptance criterion.
   • Category B: solver / driver tolerances, max iterations.
2. sim check openmdao.
3. Write .py per base/reference/workflow.md.
4. sim lint script.py.
5. sim run script.py --solver openmdao.
6. Validate JSON.