InDIA-Inversion of Subsurface Density Interface and its Application: A Machine Learning Approach to Global Optimization using Structured Gaussian Processes and Bayesian Optimization

It Bayesian Inference with Gaussian Process Regression to create a flexible and accurate Python-based inversion technique (InDIA) for predicting anisotropic subsurface layer parameters, incorporating uncertainty information. Bayesian optimization has traditionally been limited to magneto-telluric, seismic, and magnetic data. This paper introduces its first application to gravity anomaly data. By utilizing structural Gaussian Processes, the algorithm effectively enhances the accuracy of predicting inversion parameters, leveraging a diverse range of prior information. Previously, gravity data inversion techniques focused on continuous depth profiles, restricting their use to local gravity anomaly issues. In contrast, the current method estimates discretized subsurface layer depths and densities, allowing it to address local and regional gravity anomalies. This paper demonstrates the effectiveness of the proposed approach in accurately estimating inverted depths and densities, closely reflecting real-world scenarios. It addresses challenges associated with multiple inversion parameters, common in local optimization techniques such as Gradient Descent followed by ADAM-based optimizations. The superiority of this algorithm has been validated through various synthetic models featuring both constant and variable density distributions—whether lateral, vertical, or both—incorporating Gaussian noise to replicate real conditions. Furthermore, the method's efficiency in tackling multi-layered subsurface problems has been confirmed by inverting actual gravity anomaly data from the West Korea Basin in the western part of the Korean Peninsula and the Godavari Basin in the eastern part of Indian Peninsula. The resulting inverted subsurface layer depths and densities are geologically viable and align closely with previously estimated results.