An Application of Machine Learning to Predicting Shear Wave Velocity of Marine Clay

In geotechnical earthquake engineering, shear wave velocity (VS) is one of the most important parameters. Several empirical correlations have been developed to estimate VS based on in situ tests, such as P-S logging tests and piezocone penetration tests (PCPTs), and soil properties determined both offshore in a field laboratory and onshore in a geotechnical laboratory. Recently, artificial intelligence (AI), machine learning, and deep learning have been the subject of intense media hype and therefore it was deemed appropriate to develop a deep-learning algorithm (DLA) to predict VS and compare its performance with conventional empirical methods of determining VS. A database of 11 sites in the Bay of Campeche with in situ measurements of VS using P-S logging and 8 clay soil parameters, including in situ measurements of net cone resistance (qnet), water content (w), submerged unit weight (g’), void ratio (eo), undrained shear strength (su), effective overburden pressure (σ′vo), over consolidation ratio (OCR), and plasticity index (Ip), was used to train a DLA and develop empirical equations based on multiple regression analysis (MRA) to predict VS. The accuracy of the results (i.e., how close the predicted VS values are to the in situ measured VS values) illustrates the great potential DLA has to become another class of computational tool to predict VS, and DLA has the advantage that it can be continuously updated as new data become available. The DLA was found to outperform MRA when the three input variables are measurements of qnet, w, and σ′vo and when the eight clay soil parameters are used as input parameters. The results suggest deep learning has the potential to overcome prediction limitations associated with linear or polynomial trends and adjust to discrete variabilities in soil conditions. They also reveal potential applications for deep learning in reducing uncertainty in the characterization of areas with complex soils without drastically increasing the scope of a site investigation.