Conversion of Stress–strain Relationships of Wall Clay Compression Specimens With Different Geometries Using Machine Learning

This study proposes a method to convert stress–strain relationships of wall clay specimens with different geometries using statistical analysis and deep learning. Compression tests on 371 cylindrical and prismatic specimens made from Kyoto Fukakusa clay revealed systematic differences in peak strength, strain at peak, and post-peak behavior. Regression analysis proved effective for converting density and peak strength, while deep neural networks made it possible to predict strain at peak and stress–strain curves. Combining regression and DNN models provided the best overall accuracy. The proposed approach integrates heterogeneous test data, offering a unified framework for structural performance evaluation of earthen wall materials.