Machine Learning in Multi-scale Geomechanics

This project explores the application of machine learning techniques to simulate granular materials' behavior by replacing the traditional H-model with artificial neural networks (ANNs). The goal is to improve computational efficiency while maintaining the physical principles governing granular materials.

Table of Contents

1. Overview
2. Project Goals
3. Methodology
4. Results
5. Future Work
6. References

Overview

Granular materials are known for their complex behavior, often requiring intensive computational resources to simulate. This project leverages machine learning to replace the traditional H-model, a micromechanical-based constitutive model, to speed up simulations and maintain accuracy in predicting granular material behavior under different stress conditions.

Project Goals

• Replace the traditional 2D H-model framework with machine learning models.
• Generate datasets from the H-model for training and testing neural networks.
• Compare the performance of the ANN-based model with the H-model in standard geomechanical tests like isotropic compression and biaxial tests.

Methodology

The project involves several steps:

1. Data Generation: Data was generated from the analytical H-model, with incremental loading in different directions and conditions.
2. Neural Network Model: A simple feedforward neural network with 3 hidden layers of 8 nodes each was designed to predict the deformation and forces in granular materials.
3. Normalization Techniques: Various data normalization techniques (e.g., Min-Max Scaling, Robust Scaling) were used to ensure training stability.
4. Evaluation: The trained model was tested on isotropic compression and biaxial tests, comparing the results with the traditional H-model.

Results

The neural network model provided results similar to the traditional H-model in both isotropic compression and biaxial loading tests.

Key findings include:

• The machine learning model replicated the H-model behavior efficiently.
• Using robust scaling yielded the best performance for predicting the behavior of granular materials.

Future Work

• Extend to Complex Structures: Explore using more complex granular structures for simulation and testing.
• Automatic Calibration: Set up automatic calibration procedures for the machine learning model to generalize across different granular material types.