Use of ML techniques for predicting the bearing capacity of piles and its relative errors

Resumo

This work presents an application of machine learning techniques for estimating the bearing capacity of pile
foundations and the relative error from these techniques. It uses as raw data 165 load tests results associated
with SPT soundings, taken from several Brazilian regions. A dataset based on the inputs from Decourt-Quaresma
and Meyerhof semi-empirical methods was created and applied to several machine learning techniques with a
leave-one-out cross validation approach for training and testing the algorithms. Using the results obtained from
each model, the metrics RMSE and R2 were calculated through a stacking strategy. The Random Forest technique
presented the best performance for both bearing capacity (RMSE = 640,26) and relative error (R2 = 0.77) prediction
problems. The other five ML techniques performance overcame the semi-empirical methods, which obtained an
RMSE close to 900, indicating the potential of these tools. Then, the errors obtained from the predictions were
used to propose a new machine learning problem, aiming to predict the error of new examples. Although the
preliminary results were not accurate, the authors believe that the study justifies further investigations.