HBM versus TabPFN for quasi-site-specific geotechnical prediction: a cross-database evaluation

Site uniqueness is a central challenge in data-driven geotechnical prediction because information from other sites cannot be treated as direct target-site data. This study compares two contrasting approaches for quasi-site-specific prediction: the hierarchical Bayesian model (HBM), which explicitly models site-to-site variability and updates a database-calibrated prior using target-site data, and the Tabular Prior-Fitted Network (TabPFN), which uses a pre-trained foundation model and attention-based in-context prediction. Their performance is evaluated using nine real geotechnical databases covering soil-property prediction and load-test calibration. Results show that standard TabPFN is competitive when target-site data are extremely sparse, but improves only modestly as local observations increase, whereas HBM becomes increasingly advantageous through Bayesian updating. Moreover, a prior calibrated from real geotechnical databases consistently outperforms one derived solely from synthetic pre-training. Retaining simple site-level summary statistics substantially improves TabPFN’s use of increasing target-site observations, allowing it to remain broadly competitive with HBM. Overall, HBM demonstrates the value of explicit probabilistic modelling and domain-specific priors, whereas TabPFN highlights the potential of foundation models when informative context representations are provided.