ARTIFICIAL NEURAL NETWORK MODEL FOR INTEGRATING ELECTRICAL RESISTIVITY AND PENETRATION RESISTANCE OF ALLUVIAL SOILS USING TENSOR FLOW-KERAS-DRIVEN BACKPROPAGATION FEEDFORWARD ANALYSIS & HYPERPARAMETER TUNING

Abstract

This study presents an integrated approach combining Electrical Resistivity Survey (ERS) and Standard Penetration Test (SPT-N) data to improve subsurface characterization in alluvial soils, overcoming limitations of conventional drilling methods including logistical difficulties in mobilizing rotary drilling rigs measurement errors, high costs, and terrain constraints in difficult low-accessibility areas. A novel multivariate Artificial Neural Network (ANN) framework was developed using TensorFlow-Kera’s, implementing feedforward architecture with backpropagation learning. The model incorporates an innovative hyperparameter tuning protocol that systematically evaluates network depth (1-5 hidden layers) and complexity (2-10 neurons/hidden layer), identifying a 4-layer configuration provide optimal predictive accuracy (adjusted R² = 0.99, RMSE = 6.36, MAPE = 1.1%, MSLE =0.01) with effectively balancing between both underfitting and overfitting tendencies. The finalized model transforms ERS and SPT-N inputs into predictive multivariate regression equations for key geotechnical parameters estimations & foundation design analyses by applying backpropagations feedforward analysis on well trained & tested modular weight-bias matrixes of each hidden layer. This methodology advances Sustainable Development Goal 9 (SDG 9) by enabling efficient, non-invasive subsurface investigations in challenging environments (floodplains, remote areas). Specifically, it addresses Target 9.1 (resilient infrastructure development) and Target 9.4 (sustainable industrialization) through its reduced reliance on conventional drilling, demonstrating how machine learning can enhance geotechnical practice while supporting sustainable infrastructure planning.