Machine Learning Code, Dataset, and Interactive Visualisations: An Explainable AI (XAI) Framework for Particle-Modified Epoxy Fracture Energy (Gc) Prediction

Overview

This repository contains the Python machine learning code, preprocessed laboratory dataset, and interactive 3D visualisations for predicting the fracture energy (Gc) of silica-nanoparticle-modified epoxy polymers across ambient and low-temperature conditions.

Operating within a Green AI framework, this computational pipeline demonstrates how explainable non-linear regression models can deliver accurate mechanical predictions and physical insights, reducing the reliance on resource-intensive and non-recyclable laboratory testing.

Key System Variables

• Predictor 1: Particle Loading (wt%, range: 0.0 to 20.3 wt%, 20 nm silica nanoparticles)

• Predictor 2: Environmental Temperature (°C, range: -80°C to 23°C)

• Target Variable: Fracture Energy Gc (J/m²)

Preprocessing & Algorithmic Benchmarking

• Data Preprocessing: Micro-scale Core-Shell Rubber (CSR) modified epoxy samples were filtered out to eliminate high variance resulting from fundamental differences in micro-scale toughening mechanisms (such as micro-cavitation and shear-banding).

• Model Benchmarking: Four regression architectures were evaluated using 10-fold cross-validation (CV) and test split performance metrics:

  • Optimised Random Forest (Selected Model): Demonstrated the strongest generalisation stability with a Test Split R² of 0.688, 10-Fold CV R² of 0.755, MAE of 13.97 J/m², and RMSE of 18.29 J/m².

  • Gradient Boosting: Achieved high cross-validation fidelity (R² = 0.847) but exhibited overfitting on unseen data (Test Split R² = 0.440).

  • Support Vector Regression & Linear Regression: Support Vector Regression (R² = -0.094) and Linear Regression (Test Split R² = -0.220) failed to model the non-linear toughening behaviour.

Physical Validation via SHAP

SHapley Additive exPlanations (SHAP) were implemented to ensure the machine learning model aligns with physical material behaviour:

• Low-Temperature Threshold: A distinct slope change in feature attribution occurs below -40°C, identifying a critical sensitivity threshold in brittle fracture conditions.

• Toughening Plateau: SHAP dependence analysis identified a saturation point at approximately 15 wt% silica content, matching physical observations of nanoparticle debonding followed by plastic void growth.

Repository Files

1. XAI_Fracture_Prediction_Pipeline.ipynb: Executable Python notebook containing data ingestion, preprocessing routines, model benchmarking, hyperparameter optimisation, and SHAP interpretability scripts.

2. Silica_Epoxy_Dataset.csv: Cleaned experimental dataset of nano-silica modified thermosetting epoxy tested under ambient and low temperatures.

3. Interactive_3D_Fracture_Surface for Gredient Boosting: Standalone Plotly 3D response surface mapping temperature and particle content against predicted Gc, overlaid with actual laboratory measurements.

4. Interactive_3D_Fracture_Surface for Optimsed Random Forest: Standalone Plotly 3D response surface mapping temperature and particle content against predicted Gc, overlaid with actual laboratory measurements.