Understanding the current stress state of the Earth’s crust is an essential part of various geological applications, from carbon storage to fault reactivation studies. However, traditional methods are faced with significant obstacles due to the manual tuning of geomechanical properties and boundary conditions. The need for precise stress orientation data is clear, as it is pivotal for dependable geomechanical models. Manual adjustments inherent in these traditional techniques hamper the accuracy and efficiency of stress and displacement field estimations. Fortunately, a new research paper from CSIRO, Australia, tackles these issues, introducing ML-SEISMIC, a physics-informed deep neural network capable of autonomously aligning stress orientation data with an elastic model.
ML-SEISMIC’s methodology is based on applying physics-informed neural networks to linear elastic solid mechanics equations. The governing equations involve momentum balance, constitutive relationships, and small strain definitions. The neural network optimizes stress field eigenvalues concerning stress orientations, thus providing a comprehensive understanding of the stress and displacement fields. The application of ML-SEISMIC to Australia serves as a case study, revealing its ability to autonomously retrieve displacement patterns, stress tensors, and material properties. The method is successful in overcoming the shortcomings of traditional approaches, offering a reliable interpolation framework. Additionally, it utilizes Global Navigation Satellite Systems (GNSS) observations to revisit large-scale averaged stress orientations and identify areas of inconsistency. The results demonstrate the adaptability of the approach across various scales, from crystallographic investigations to continental-scale analyses.
In short, ML-SEISMIC is a revolutionary solution in geological investigations. This physics-informed neural network offers a streamlined and powerful process, nearly eliminating the need for explicit boundary condition inputs. It is flexible enough to be applied to different scales, all while relying on accurate GNSS observations. ML-SEISMIC is sure to be a game-changer in the ever-evolving landscape of scientific inquiries, providing a versatile and powerful tool to gain further insights into Earth’s dynamic processes.
