Optimal Experimental Design with Neural Network Surrogates

Authors: Joshua Stuckner, Matt Piekenbrock

This code was used in the following papers:

• Joshua Stuckner, Matthew Piekenbrock, Steven M Arnold, Trenton M Ricks. (2021) Optimal Experimental Design With Fast Neural Network Surrogate Models To be published in Computational Materials Science
• Joshua Stuckner, Matthew Piekenbrock, Steven M Arnold, Trenton M Ricks. (2021) Optimal Experimental Design With Fast Neural Network Surrogate Models. NASA Technical Reports Server, TM-20205003868
• Arnold, S. M., Piekenbrock, M., Ricks, T. M., & Stuckner, J. (2020). Multiscale Analysis of Composites Using Surrogate Modeling and Information Optimal Designs. In AIAA Scitech 2020 Forum (p. 1863).

How to use

The entry point to run the four experiments in the Computational Materials Science paper are in the vignettes folder. Each experiment has its own notebook: PMC_class.Rmd, MMC_class.Rmd, CMC_class.Rmd, VF_experiment.Rmd. The data and trained models for these experiments are in the data folder.

The VF_experiment will be used as an example.

Step 1 - Create training data (lines 30 - 121)

This step cannot be performed without MAC/GMC, a physics based composite modeling software. The parsed data from this step is included in the data folder and this step may be skipped.

Step 2 - Load and clean the data (lines 124 - 213)

Step 3 - Train the neural network surrogate (lines 214 - 372)

This step can be skipped by loading the trained models directly from the data folder. This step ended up being replaced for the VF experiment by the Hyperparameter optimization step.

Step 3b - Hyperparameter optimization (lines 375 - 523)

This step can be skipped by loading the trained models directly from the data folder.

Step 4 - Perform OED (lines 525 - 801)