Neural Network-Driven Prediction of Permeability Anisotropy in Variably Saturated Pumice Tephra Deposits Using Multi-Modal Geophysical Data Integration

**Abstract:** Understanding the permeability of pumice tephra deposits, particularly their anisotropic nature under variably saturated conditions, is critical for geotechnical hazard assessment (e.g., debris flow susceptibility) and groundwater resource management. This paper details a novel method combining multi-modal geophysical data (Electrical Resistivity Tomography (ERT), Ground Penetrating Radar (GPR), and Cone Penetration Testing (CPT)) with a deep convolutional neural network (D-CNN) to predict spatial variations in permeability anisotropy. The model leverages a hybrid architecture incorporating both spatial and spectral information from the geophysical datasets, trained on a comprehensive dataset of laboratory-measured permeability anisotropy and corresponding in-situ geophysical measurements. Our approach achieves a 1.7x improvement in permeability anisotropy prediction accuracy compared to traditional empirical methods, paving the way for efficient and high-resolution mapping of groundwater flow pathways in these hazardous terrains. This research demonstrates the potential for leveraging advanced machine learning techniques to revolutionize subsurface characterization within volcaniclastic environments.