Neural transfer learning for soil liquefaction tests

Soil liquefaction is one of the most disastrous sides of earthquakes which can cause severe damage to structures, infrastructures, and individuals’ lives. Therefore, establishing new and advanced models to predict the liquefaction potential of soil is an inevitable task to evaluate sites quality. Considering the shortage of data for some soil liquefaction tests and the independence among the models built for each soil liquefaction test, it is difficult to provide highly accurate models. In this paper, a new approach is proposed to evaluate soil liquefaction based on transfer learning to generalize over the soil liquefaction tests or measurements (Vs, CPT, DPT, and SPT). Firstly, a pre-trained model is built using the artificial neural network (ANN) and the comprehensive shear-wave velocity (Vs) test dataset which contains a total of 1069 cases from different sources. Secondly, the developed pre-trained model is evaluated using classification evaluation metrics with a focus on the recall metric, because there is a high cost associated with false negative cases (actual liquefied cases classed as non-liquefied), with a recall value of 90% and an accuracy value of 76%. Finally, the model is validated by comparing its predictions on other less-resourced tests (CPT, DPT, and SPT) with some well-known models in each test by using only 20% of the dataset to fine-tune the developed pre-trained model. Compared with DPT, CPT, and SPT models on recall metric, our model is the most optimal one with recall values of 92%, 97%, and 89%, respectively. The results show that applying a pre-trained model to predict soil liquefaction potential using other soil liquefaction tests is an effective approach instead of training the models separately from scratch and/or testing with limited data. Moreover, other metrics indicate that the proposed approach outperforms existing classical machine learning approaches and presents the state-of-the-art in terms of transfer learning.