Lessons learned from training an MLP model using layered feature sets for estimating pile bearing capacity

Purpose This study aims to investigate how input representation and training-set size influence the performance of multilayer perceptrons (MLP) for predicting axial pile capacity. Empirical procedures often show wide deviations from full-scale load tests, motivating data-driven approaches that better capture soil and geometry variability. Design/methodology/approach A database of 546 axial load tests is used to evaluate 12 feature representations ranging from geometry-only inputs to averaged soil descriptors and layered soil profiles sampled along the pile shaft. For each feature set, an MLP is tuned using 500 Optuna trials under three training sizes: one-third, two-thirds and the full data set. Findings Layered inputs generally outperform geometry alone, but averaged soil properties often perform similarly. The combination of geometry with layered soil type, total unit weight and SPT-N provides the most consistent results. Training size is the dominant factor, with large gains from one-third to two-thirds of the data and smaller gains thereafter. Practical implications Large geotechnical data sets can support reliable ML models even when many soil parameters are missing or inferred. Averaged soil information performs nearly as well as layered profiles, reducing the need for complete stratigraphic data. Practitioners can therefore use broad but imperfect databases to develop effective prediction tools. Originality/value This study provides a comprehensive assessment of feature design and data sufficiency for neural-network-based pile capacity prediction. It demonstrates that while representation matters, training-set size governs most achievable accuracy and clarifies when layered inputs justify added complexity.