Deep learning-based calibration of resistance factors for pile groups with load tests

Abstract Resistance factors for pile groups are typically derived using empirical methods that do not directly account for system redundancy and overlook the correlation between individual piles, which are inherently influenced by the spatial variability of soils. While rigorous three-dimensional (3D) random finite difference (RFD) or random finite element (RFE) analyses could potentially address these issues, they are constrained by significant computational demands. Therefore, this paper proposes a deep learning-based approach for calibrating resistance factors for pile groups with individual pile load tests. Specifically, a surrogate model based on a convolutional neural network (CNN) is proposed, which is trained and validated using the database generated by RFD analyses. The trained model is further used to derive pile resistances in spatially variable soils. Finally, the resistance factors are calibrated by counting and conditional probability based on the outcomes of load test results. The proposed approach is demonstrated using a pile group example. Results show that the proposed approach effectively captures the impacts of load test results and their corresponding locations, as well as the spatial variability of soil properties, on resistance factors.