Yeh, Fu-HsuanFu-HsuanYehLiang, WeiWeiLiangYU-NING GEHsiao, Cheng-HsiCheng-HsiHsiao2026-03-122026-03-12202619648189https://www.scopus.com/record/display.uri?eid=2-s2.0-105028367004&origin=resultslisthttps://scholars.lib.ntu.edu.tw/handle/123456789/736227Due to spatial variability in soil properties, the random finite element method (RFEM), which incorporates the random field concept, has been introduced to consider slope uncertainty and reliability. The RFEM requires a quantitative evaluation of a slope failure probability, which is time-consuming, so machine learning techniques can serve as an alternative way to predict slope failure in a timely fashion. Recently, a convolutional neural network (CNN) model has been a suitable method that can be trained by inputting spatial variability relationships to address the problem of slope failure evaluation. This study developed a pre-trained CNN model using RFEM simulation results for 30° and 50° slope inclinations, considering spatial relationships and soil variability using the coefficient of variation and scale of fluctuation. The model is then tested on unseen 40° slopes to evaluate its generalisation capability. The results indicate that both the basic shallow CNN and ResNet-enhanced models can make accurate predictions, with the ResNet-6 configuration (a six-layer model) demonstrating the best performance. The integration of ResNet-6 improved the model’s ability to generalise across different slope geometries in terms of safety factor prediction.falseand scale of fluctuationcoefficient of variationconvolutional neural networkrandom finite element methodresidual neural networkSafety factorjournal article10.1080/19648189.2026.26154502-s2.0-105028367004