Explainable surface displacement prediction using causal feature selection and Kolmogorov–Arnold Networks

Potential failures from unstable rock slopes and landslides pose a significant risk to human beings and infrastructure, making the understanding and prediction of surface displacement critically important. We introduce a novel framework leveraging the most recent advances in machine learning and aiming at comprehensively understanding the displacement behaviour of unstable slopes. The construction of causal graphs connecting available external and internal slope parameters allows identifying the underlying displacement drivers and extracting valuable information to train predictive models. Our approach utilizes a Kolmogorov-Arnold Network (KAN)-based architecture to enhance prediction capability. We applied the displacement prediction models to datasets from three slope instabilities characterised by distinct geological and geographical settings and varying levels of monitoring data availability, demonstrating that our end-to-end framework is widely applicable. Our method not only outperforms almost all recent and standard machine learning approaches in terms of prediction accuracy, but also offers easy interpretability for the end user. The employment of accurate and at the same time easy-to-understand models is a significant advancement, overcoming adoption issues of complex Artificial Neural Networks. The inherent model transparency makes this approach relevant for a wide range of geoscientists, with the majority being non-experts in machine learning, ultimately contributing to better understanding of slope behaviour and thus better mitigation strategies and risk management.