Neural network architecture and training data for prediction of porous material mechanical properties based on their microstructure

This repository contains data used for predicting the mechanical response of porous natural materials based on their microstructural features. CT scans of porous natural materials were segmented using ImageJ (and the custom ImageJ script SegmentJ: https://github.com/JPELO/SegmentJ), and microstructural geometric features such as volume density, surface area density, and asymmetry were extracted from the scans. Mesh .STL files were also created using ImageJ for each sub-scan. The open source parallelizable finite element software MOOSE was used to simulate uniaxial monotonic compression testing and produce a stress-strain response for each sub-scan. Additionally, the mesh files were used to 3D print physical microstructures and were tested experimentally in uniaxial monotonic compression at a rate of 5mm/min. The material used to print the structures was Formlabs Grey Pro with a Formlabs Form2 printer. MATLAB was then used to create and train a neural network capable of predicting the mechanical stress-strain curve of a porous microstructure based on its geometric features.