OpenSees ML, Sweep to Regression

An end-to-end machine learning workflow on DesignSafe. A PyLauncher sweep runs a 2D cantilever pushover in OpenSeesPy across a 75-point parameter grid, and a post-script fits a linear regression that recovers the physics from the results, all in one python-s3 job.

For PyLauncher basics, see PyLauncher Parameter Sweeps. For the app settings used here, see the Generic Python App.

The period equation as ground truth¶

The natural period of a cantilever with tip mass follows

T = 2π · sqrt(M · L³ / (3 · E · I))

so in log space the exponents are known exactly, 0.5 on mass, 1.5 on length, and -0.5 on modulus. The workflow sweeps NODAL_MASS x LCOL x EMOD (5 x 5 x 3 = 75 OpenSeesPy pushover runs), aggregates each task’s metrics.json, and fits log T against the log features with scikit-learn. Because the physics is exact, the regression must report coef ≈ [0.5, 1.5, -0.5] with R² = 1.0, which makes the example self-checking. Wrong plumbing cannot produce the right coefficients.

How one job does all of it¶

Each stage of the python-s3 job does one piece of the work.

The sweep definition and submission are three notebook cells.

sweep = {
    "NODAL_MASS": [4.19, 4.39, 4.59, 4.79, 4.99],
    "LCOL": [100, 200, 300, 400, 500],
    "EMOD": [3600, 4000, 5000],
}

commands = ds.jobs.parametric_sweep.generate(command, sweep, str(input_dir))
job = ds.jobs.parametric_sweep.submit(...)

The placeholder is named EMOD rather than E on purpose. Token-style placeholders match anywhere a word appears, so a single-letter key like E also rewrites the --E flag in every generated command.

Results¶

The archived job contains ml_results/ with the fitted model as JSON, per-task predictions as CSV, and a diagnostics figure (histograms, predicted versus truth, residuals, coefficients). The notebook downloads and displays them.

job.download_output(
    "inputDirectory/ml_results/opensees_ml_diagnostics.png",
    "opensees_ml_diagnostics.png",
)

On Stampede3, the 75 tasks take about 42 seconds of task time, and the regression recovers [0.500, 1.500, -0.500] with R² = 1.0.

The same study as a DAG workflow¶

DS_OpenSees_ML_Workflow_DAG.ipynb splits the pipeline into two jobs, a 48-core sweep and a one-core training task, connected as an explicit graph with dapi.workflows. dapi points the training job’s input at the sweep’s archive through an output reference and resolves it to the real path before submitting, retraining costs one core instead of a resweep, and the run streams live per-task progress in the notebook. The Workflows guide covers the API.

Files¶

The example also runs locally without Tapis. The README shows the two-task local dry run.