STUDY ON SOIL MIXTURE STRENGTH DEVELOPMENT FORECASTING BASED ON BP NEURAL NETWORK AND SVM

Abstract

The strength value is a core indicator in the field of soil stabilization, so evaluating the strength development is a common problem in geotechnical engineering. Compared with the traditional linear fitting method, Machine Learning applied in the engineering field is a more effective and accurate method for fitting and predicting data. The aim of this research was to explore whether the unconfined compressive strength values of soil mixtures can be accurately predicted by machine learning model. Back Propagation Neural Network (BPNN) and Support-vector Machines (SVM) are classical algorithms in the Machine Learning field, and both algorithms will be tested in this research. On the other hand, to improve the precision rate, this research uses an empirical formula to determine the quantity of hidden layers of the Back Propagation model and uses grid searching to find the best penalty factor value and size of mapped dimensions of the SVM. The results showed that both Machine Learning algorithms have higher accuracy, but that SVM has a better performance compared to BPNN: the average R2 value of SVM is 0.94426; the average R2 value of BP NN is 0.94426. It provides a new statistical method for soil performance research in the future.

Keywords: Machine Learning, BP Neural Network, Support-vector Machines, Soil improvement, Unconfined compressive strength