A physics-guided neural network for reliable slope stability assessment under data scarcity

Abstract

Slope stability prediction is a critical task in geotechnical engineering, yet regional-scale applications are often prohibitive due to the high computational costs of existing physics-based methods. While data-driven surrogate models offer a computationally efficient alternative, their predictive capabilities are usually hindered by the limited availability of high-quality real-world slope stability datasets. To address this challenge, this study proposes a novel physics-guided neural network (PGNN) framework for slope stability prediction. The novelty of this framework lies in that (1) we integrate classical stability charts as physical constraints via a staged-loss training schedule, and (2) we propose a feature-importance-based clustering strategy for data partitioning to ensure representative coverage of dominant geotechnical variables in both training and validation subsets to further enhance model robustness. The proposed framework is developed and evaluated using a compiled real-world slope dataset, with data consistency ensured through a unified Factor of Safety (FoS) recomputation. Results demonstrate that the proposed PGNN surrogate reduces initialization sensitivity and suppresses extreme prediction outliers in distribution-shifted scenarios compared to purely data-driven baselines. In a regional-scale application across an 8547 km 2 topographically complex area in Washington State, the trained surrogate provided rapid post-training inference and achieved high classification agreement with Bishop-based reference calculations (F1-score ≈ 0.89); relative to the single-threaded Python Bishop workflow used in this study, this corresponded to a 15,612-fold reduction in evaluation time. This study provides an accurate and efficient surrogate for rapid Bishop-based regional slope-stability screening.