Explainable AI-Based Enhanced Classification of Gravitational Wave Anomalies Using Vision Transformers

Abstract

Gravitational wave Observatories like LIGO detect both genuine gravitational wave signals and various forms of noise known as "glitches". Our research tackles this challenge with a machine learning pipeline enhanced by explainable AI (XAI). The proposed model begins by processing signal spectrograms with Vision Transformers (ViT) and Swin Transformer to extract detailed visual patterns. Then augment these patterns with numerical features (e.g., event duration, frequency) to create a comprehensive dataset. Several models—XGBoost, Support Vector Machine, Artificial Neural Network, Random Forest, and Decision Tree—are evaluated on the Gravity Spy dataset. The results show XGBoost, combined with ViT-derived features, attains 94.40% accuracy, outperforming other approaches. To build trust in the model’s decisions, we apply explainability tools like SHAP and LIME, which pinpoint the specific visual and numerical elements that drive each prediction. This high level of accuracy, paired with transparent insights, not only refines glitch detection but also boosts researchers’ confidence in gravitational wave studies.