A ML Approach for Reliability-Based Design of Vertically Loaded Piles Using Sparse CPT Data

Cone penetration tests (CPT) provide tip resistance (qt) and sleeve friction (fs) profiles for pile design, but data are often sparse due to cost and access, creating uncertainty at untested locations. This study proposes a novel ML-based method for CPT probabilistic prediction at untested locations in layered grounds, incorporating CPT and soil type profiles. The method consists of: 1) soil classification at CPT locations with no borehole data available, 2) stratigraphy interpolation in between CPT locations, and 3) CPT probabilistic predictions for pile capacity evaluation using spatial coordinates and soil type as inputs. Simulated qt and fs data were generated using Random Field with five soil types and a defined correlation structure. Results showed that including soil type in addition to coordinates improved prediction accuracy for both qt and fs, with higher R² and lower errors compared to the coordinate-only model. Using the distribution of predicted CPT, the Unified CPT-based axial pile capacity method was applied to an assumed closed-ended pile, showing ∼30% reduction in CoV and ∼60% reduction in error at several untested locations. Overall, findings from simulated data indicate the potential of this ML-based approach; however, further validation is needed using real-world field measurements.