Interpretation of subsurface stratigraphic variations from limited boreholes using Dual Bi-LSTM

The determination of stratigraphic delineation is the core of geotechnical design, significantly influencing the safety and serviceability of infrastructures. Accurately inferring subsurface stratigraphic variations from sparse borehole data still presents a considerable challenge due to the complicated spatial correlation of soils. In this paper, the dual bidirectional long short-term memory (Bi-LSTM) neural network is developed for accurate stratigraphic delineation by using limited boreholes. Based on a novel cross-shaped sampling system, the Dual Bi-LSTM efficiently captures intricate spatial dependencies. Moreover, numerical and one-hot encoding methods are compared to explore different ways of representing stratigraphic features. The proposed model is validated and implemented in three practical projects collected from Australia, Hong Kong, and the Netherlands, respectively, compared with the Markov random field (MRF) and IC-XGBoost method. Furthermore, the effects of borehole density, neighborhood scale, and sampling scheme are investigated based on a nonlinear and non-homogeneous synthetic case. The proposed model achieves an accuracy of 60.35% in boundary predictions of the Australia case, surpassing the MRF and the IC-XGBoost model by around 6% and 23%, respectively. The proposed Dual Bi-LSTM is highlighted to provide a user-friendly alternative for involving an accurate and reasonable soil profile using sparse site-specific boreholes.