Predição da resistência à compressão simples de solos estabilizados utilizando aprendizado de máquina automatizado

Abstract

Soil stabilization using various types of additives, most commonly chemical additives such as lime and cement, has been used for centuries to improve soil properties in geotechnical engineering projects. Results from the unconfined compressive strength (UCS) test are used to monitor and evaluate the increase in shear strength of stabilized soils. However, the large number of tests required on representative samples of stabilized soil makes experimental evaluations time-consuming and costly. The objective of this study is to apply and compare five automated machine learning (AutoML) systems—AutoGluon, AutoKeras, FLAML, H2O, and TPOT—for predicting the UCS of stabilized soils. Unlike previous approaches that automated only hyperparameter tuning, this research proposes a fully automated workflow, including preprocessing, feature selection, and algorithm selection. The methodology consisted of analyzing ten experimental datasets under varied geotechnical conditions, totaling 1,745 samples from multiple countries and covering stabilizing agents such as cement, lime, and agricultural residues. Performance was evaluated using regression metrics and a composite Performance Index (ID). The results indicate that the AutoML systems consistently outperformed the models reported in the original literature for each dataset. The FLAML system demonstrated the most stable and efficient overall performance, achieving the highest average ID (0.78). Feature importance analysis revealed that the models successfully captured fundamental physical and chemical mechanisms, prioritizing stabilizer content and compaction parameters. It is concluded that AutoML is a robust tool for geotechnical engineering, enabling the rapid development of predictive models and reducing reliance on extensive experimental procedures, while also mitigating the need for specialized expertise in data science. Nevertheless, specifically optimized manual models may still achieve marginally higher levels of accuracy.