Deep Learning-based Microseismic Source Location with Joint Constraints of Source Imaging and Traveltime Residuals

Summary Microseismic source location is essential for seismic monitoring and subsurface resource exploitation. Both traveltime inversion and waveform stacking methods suffer from limited accuracy when processing low signal-to-noise ratio (SNR) data under complex velocity models. Existing deep learning approaches mainly employ purely data-driven strategies without physical constraints, exhibiting limited capability to suppress large and unexpected location errors. We propose a physics-constrained deep learning method for microseismic source location that integrates the physical principles of cross-correlation stacking (CCS) imaging into network training. The method incorporates a joint loss function combining source imaging quality loss and traveltime consistency loss, with a Pareto dynamic weighting strategy to balance different loss components. Synthetic experiments on the Marmousi velocity model demonstrate that the joint-constrained method reduces the mean absolute error (MAE) from 34.09 m to 27.91 m compared to the purely data-driven approach. The maximum error decreases from 280.18 m to 130.38 m, a 53.5% reduction, demonstrating effective suppression of large location errors. The trained network achieves single-event imaging prediction in 0.04 s, providing a 75-fold speedup over the 3 s required by conventional CCS. The proposed method shows great potential in near-real-time microseismic monitoring with dense arrays.