Physics-guided multimodal deep learning reveals determinants of rock brittleness across scales

Abstract

This study proposes a physics-guided multimodal deep learning method for predicting rock brittleness from integrated microstructural and compositional data. The multimodal framework was constructed by integrating scanning electron microscopy (SEM), energy dispersive spectroscopy (EDS), and X-ray diffraction (XRD) data. Vision Transformers were employed for microstructural image analysis, while specialized neural networks were designed for compositional data processing. Feature fusion was achieved through attention-based mechanisms to maintain physical interpretability. The multimodal approach significantly outperformed single-modality methods (R² = 0.995, RMSE = 0.065, MAE = 0.051). Feature correlation analysis was conducted to identify key determinants, revealing Si/Al ratio, quartz content, and porosity as primary controlling factors. Laboratory validation was performed using three representative rock types (granite, red sandstone, and green sandstone) through uniaxial compression tests. The model predictions showed excellent agreement with experimental brittleness indices, demonstrating superior discrimination capability with expanded dynamic range compared to traditional methods (B₁, B₂). Fragmentation analysis using mean particle size (dm) provided additional validation, confirming the trend that predicted brittleness increases as post-failure particle size decreases. Granite samples exhibited the highest brittleness (Bpred = 3.62–4.87) and finest fragmentation (dm = 16.8–19.5 mm), while green sandstone showed the lowest brittleness (Bpred = 0.74–0.95) and coarsest fragmentation (dm = 21.3–30.7 mm). The results demonstrate that the proposed multimodal deep learning framework effectively captures the complex relationships between microstructural features and rock brittleness, offering significant potential for accurate brittleness prediction and enhanced understanding of rock failure mechanisms.