A bayesian neural network framework for CPT prediction at sparse sites
Document Type
Article
Publication Date
2-1-2026
Abstract
Geotechnical site characterization using in-situ tests, such as cone penetration tests (CPTs), is essential for foundation design but is often limited by sparse spatial coverage, hindering accurate soil variability assessment. This study benchmarks six prediction techniques, Bayesian compressive sampling with Markov chain Monte Carlo (BCS_MCMC), Bayesian neural network (BNN), genetic algorithm (GA), gene expression programming (GEP), empirical Bayesian kriging (EBK), and inverse distance weighting (IDW), to predict corrected cone tip resistance (qt) at untested locations across ten Louisiana sites. The performance of these techniques is evaluated using the root mean square error (RMSE), mean absolute percentage error (MAPE), mean bias factor (λ), coefficient of efficiency (COE), coefficient of variation (COV), and a unified Performance Index (PI) analysis. Results show that BNN, EBK, and IDW consistently achieve higher accuracy, stability, and minimal bias; whereas GA, GEP, and BCS_MCMC exhibited larger errors than BNN/EBK/IDW when validated against measured qt profiles. Prediction quality depends strongly on CPT layout, with favorable accuracy at minimum spacing near ∼100 ft and distribution indices between ∼0.05–0.10. The proposed BNN architecture is implemented in the CPT Site Variability Suite (CSVS), a MATLAB tool developed by the authors that automates data processing, interpolation, visualization, and downstream analyses (e.g., variogram derivation and LRFD workflows), all within a single platform. This integrated pipeline enhances reproducibility and supports data-driven foundation design in geotechnical site investigations. Findings pertain to the Louisiana dataset examined and provide a transferable workflow that should be validated for other geologic settings.
Publication Source (Journal or Book title)
Transportation Geotechnics
Recommended Citation
Moradi, P., Al-Harahsheh, Y., Nobahar, M., & Abu-Farsakh, M. (2026). A bayesian neural network framework for CPT prediction at sparse sites. Transportation Geotechnics, 57 https://doi.org/10.1016/j.trgeo.2025.101837