Estimativa do índice de compressão de solos argilosos moles: uma abordagem de aprendizado de máquina automatizado

Abstract

Settlement is a phenomenon of great interest in geotechnical engineering, related to the vertical deformation of soils and important in preventing structural issues in constructions. To perform the calculation, a property called compression index is required, which can be obtained through laboratory tests. However, these procedures can be costly in terms of time, equipment and specialized labor. On the other hand, in recent decades, literature has presented empirical equations and machine learning models to estimate soil properties and results from geotechnical tests. Furthermore, in recent years, automated machine learning (AutoML) has emerged as an alternative to train and stack several models with simpler procedures. Therefore, this study aims to develop AutoML models for the prediction of soil compression indexes and to compare their performances. In order to do so, four databases with soil samples from different places were selected. These sets underwent data preprocessing and were then applied to three models: AutoGluon, FLAML and H2O. The evaluation metrics used were R2 , RMSE and execution time. In the end, it was found that the three models presented similar R2 and RMSE values, with FLAML showing the most satisfactory values in most cases. H2O produced similar results but with a considerably shorter execution time. It was concluded that, despite having different strategies, the three models perform similarly under the same conditions.