Practical Application of Machine Learning to Geotechnical Correlations

Abstract

Abstract Correlations are prevalent in the geotechnical engineering practice. This paper presents a practical machine learning approach for establishing improved geotechnical correlations. In the realm of machine learning, multiple input variables of significance can be readily and coherently incorporated into the model. This approach is illustrated in the context of correlation between undrained shear strength and the combination of depth, water content, liquid limit and plastic limit. The methodology is presented in a general form to facilitate adaptation to other geotechnical correlations. The machine learning model is created in TensorFlow which is an open source machine learning platform. A dataset which consists of 1013 groups of undrained shear strength from undrained-unconsolidated triaxial test, depth of sample, water content, liquid limit and plastic limit data are compiled. 70 % of the randomised dataset are used for training and validating the model while the remaining 30 % are used to test the performance of the model. The machine learning model consists of an artificial neural network model with one input layer, four hidden layers and one output layer. The training takes several minutes on a laptop equipped with a Graphics Processing Unit. The classical approach would have been to correlate the undrained shear strength to liquidity index which is a combined parameter computed from water content, liquid limit and plastic limit. By comparing the predictions of the machine learning model on the test dataset against those computed based on dataset-specific correlation between undrained shear strength and liquidity index, it is evident that a significantly improved correlation is obtained using the machine learning approach. Due to the intricacy of multi-dimensional regression analysis, classical geotechnical correlations are typically determined by method of curve-fitting with just a single independent variable. Soil, by nature, is a complex material that is characterised by multiple index properties. Therefore, it is intuitive that geotechnical correlations can be improved by incorporating multiple classification properties or measurements. The machine learning approach described in this paper, which is implemented using an open source platform and readily accessible to industry practitioners, alleviates the legacy limitation of a single independent variable leading to improved geotechnical correlations. More importantly, this machine learning approach fits in perfectly with the digitalisation initiative increasing embraced by the oil and gas industry to improve on safety, efficiency and profitability.