Machine learning-based mapping of the vertical scale of fluctuation of spatially varying soils by assimilating the site-specific SPT profiles

Modeling of inherent spatial variability of heterogenous ground is generally performed using the random field theory characterized by the mean value, correlation distance, and scale of fluctuation (SOF). SOF is a parameter for describing the spatial heterogeneity of soil, used in random field modeling of geotechnical problems. However, estimating the SOF, considering the site investigation data alone is challenging due to the apparent limited information. In this regard, this paper aims to develop a convolutional neural network-based approach to estimate the vertical SOF of heterogenous ground by assimilating the SPT borehole profiles in the sense of reducing the sparsity of data points in the SPT data by refining the interval of sampling points. This machine learning model is then employed in an actual site project of a 7 km long bridge over the Brahmaputra River between Majuli and Jorhat, India. To validate the estimations, a comprehensive comparative study is conducted to evaluate the applicability of modern approaches over traditional methods. Moreover, a Bayesian statistical approach is also incorporated in this study to increase the reliability of the frequently used Autocorrelation functions (ACFs). The vertical SOF values obtained in this study fall within a reasonable range and hence proposed methodology shows a good agreement in capturing the correlation trend of SPT based data.