pyoptmat: statistical inference for material models

pyoptmat is a package for calibrating statistical material models to data. The package is based on pytorch and pyro and provides a framework for using machine-learning techniques to calibrate deterministic and statistical models against experimental data.

A “material model” is mathematically a parameterized system of ordinary differential equations which, integrated through the experimental conditions, returns some simulated output that can be compared to the test measurements. pyoptmat uses Bayesian inference with the pyro package to find statistical distributions of the model parameters to explain the variation in the experimental data.

As an example, consider a collection of tension test data on several samples of a material. The test measurements have some variation caused by manufacturing variability and uncertainty in the experimental controls and measurements.

pyoptmat aims to make training a statistical model to capture these variations easy. The image shows the results of training a simple material model to the test data. The trained statistical model captures the variability in the experimental data and can then be used to translate this uncertainty to models of engineering components. Transferring uncertainty quantified in experimental measurements to predictions of uncertainty in engineering applications is the main reason pyoptmat was developed.

The software is provided under an MIT license. Full documentation is available here.