Predicting rapid impact compaction of soil using a parallel transformer and long short-term memory architecture for sequential soil profile encoding

Abstract

This study presents an advanced deep learning approach for predicting the effectiveness of Rapid Impact Compaction (RIC). The model integrates the focused attention mechanisms of transformer architectures with the sequential data processing capabilities of Long Short-Term Memory (LSTM) networks. Input parameters include the initial soil profile and feature vectors representing the soil's initial state, applied compaction effort, and compaction hammer energy. Utilizing an encoder-decoder framework, the model encodes soil profile information at various depths into tokens, which are subsequently decoded to predict the resulting ground improvement. An ablation study was conducted to assess the significance of each model component. The model's predictive accuracy was validated using field test data, demonstrating a strong correlation with observed outcomes (mean absolute error of 0.42 for test data). Shapley value analysis of the trained model revealed that compaction effort exerted the highest influence on predictions, followed by fine content and fill thickness. The model architecture also demonstrated successful application to alternative RIC case studies, indicating potential generalizability. Furthermore, the model's capability to simulate hypothetical scenarios with varying compaction efforts provides valuable insights for strategic planning and optimization of RIC project designs.