Neural Networks for Enhanced Pile Design Coefficients in Effective Stress Beta Method-Article

Article

Pdf Summary

The study evaluates whether neural networks can improve selection of the β and Nt coefficients used in the effective-stress Beta method for designing driven piles. The authors compared FHWA-calculated capacities with measured load-test results from the Deep Foundation Load Test Database (DFLTD V2), focusing primarily on steel H-piles.<br /><br />For 59 projects, capacity estimates using FHWA guidance varied substantially from measured values: the calculated-to-measured capacity ratio ranged from 0.2 to 5.6. The researchers back-calculated β and Nt from load-test data and found that these coefficients often differed from FHWA recommendations. They then trained and tested machine-learning models using soil and pile properties as inputs. The best models were BetaSPTNet, a single-hidden-layer neural network for β, and NtSPTNet, a two-hidden-layer network for Nt.<br /><br />On unseen test data, BetaSPTNet achieved an R² of 0.72, while NtSPTNet achieved an R² of 0.81. Applying the predicted coefficients improved capacity estimates in an external validation of a closed-end steel pipe pile: the neural-network estimate was closer to the measured capacity than the estimate based on FHWA coefficients. Feature-importance analysis identified effective soil unit weight as most influential for β predictions and friction angle as most influential for Nt predictions.<br /><br />The authors propose incorporating these neural-network predictions into the conventional Beta-method design process. They note that the models were developed from a relatively limited database, including H-pile projects, and recommend expanding the training data and testing performance under more varied conditions. Overall, the findings suggest that neural networks may improve coefficient selection and reduce uncertainty in driven-pile capacity estimates, but broader validation is needed before general application.

Keywords

neural networks

driven piles

Beta method

effective stress

pile capacity

FHWA guidance

load test database

steel H-piles

machine learning

soil properties