Using Machine Learning (ML) Method to Improve Correlations between SPT and TCP Data

In situ testing is an essential part of field investigation, as it provides substrata information that lab testing cannot. The Standard Penetration Test (SPT) is one of the most popular in situ methods to indirectly provide soil properties. However, some of the Southwest and Midwest states predominantly conduct Texas Cone Penetrometer (TCP) tests for their practice. Extensive effort has been made to establish a correlation between SPT and TCP blow counts so existing correlations between TCP/SPT blow counts and soil properties such as friction angle, cohesion, and modulus can be shared. However, the outcomes were unsatisfactory. For example, the R2 (coefficient of determination) of TCP and SPT correlation always fell below 0.5. The major reason for being unable to develop a strong correlation between TCP and SPT is that only TCP and SPT blow counts were accounted. This study utilizes machine learning (ML) approach to include soil gradation, Atterberg limits, and depth as independent impact factors to develop correlation between TCP and SPT. Both linear and non-linear relationships are considered when these impact factors featuring soil and site characteristics are included. Based on this study, the R2 can be increased to greater than 0.8, compared with 0.05 when only blow counts are considered. The sensitivity of each impact factor on the correlation between TCP and SPT is investigated by excluding one factor each time from the ML algorithm. The results indicate that once a factor is excluded from the model, the R2 decreases. This study clearly indicates that many factors impact the TCP and SPT correlation, and they should be considered. This study also gives us a new horizon for future correlation: unlike simple one-to-one correlation of TCP and SPT, ML provides us with an approach to reveal comprehensive correlations by including many factors.