A practical machine learning approach to predict soil layering for surface wave data inversion: Framework and case study

Abstract

Traditional geotechnical site investigations often depend on invasive borings, which require significant time and resources. In contrast, surface wave methods gather data without disturbing the soil and are faster and more cost‐effective, allowing investigation of more extensive areas. However, these techniques involve an inversion process which necessitates significant computational effort and often leads to multiple solutions. The convergence of the inversion process and the accuracy of its output heavily rely on the initial assumptions about soil layering, including the number of layers, shear‐wave velocity and density. In this study, a machine learning (ML) approach was developed to refine the estimation of the most appropriate number of soil layers for a model. Nine hundred thousand synthetic shear‐wave velocity soil profiles were generated, each with a total thickness ranging from 20 m to 60 m. The number of layers ranges between two and seven, with each layer having a randomly determined thickness and shear‐wave velocity, then converted into Rayleigh and Love dispersion curves and horizontal‐to‐vertical spectral ratio (HVSR) curves for ML training. Three separate tail‐masked gated recurrent units with decay classifiers were then trained for each data type. To determine the best‐performing models, various network architectures with different optimizers were tested, and their confusion matrices and test accuracies were evaluated. The selected models were then used to produce a probability vector indicating the most likely number of layers for each individual Rayleigh dispersion curve, Love dispersion curve or HVSR curve. Based on these probability vectors, the final layer count prediction was made. These models demonstrate satisfactory accuracy when validated against field data, offering a practical way to refine initial assumptions in the inversion process for prediction of the number of model soil layers for the inversion process. All three classifiers are bundled in the soil layer estimator GUI, which visualizes input curves, enables adjustment of per‐model weights and exports Dinver‐compatible parameter files for the inversion setup. This approach provides a versatile tool for improving both the speed and precision of the inversion process for surface wave investigations.