MRGAN: a multi-resolution residual generative adversarial network with wavelet decomposition and phase-aware loss for microseismic waveform denoising

Microseismic monitoring plays a critical role in hydraulic-fracturing evaluation, enhanced geothermal system (EGS) development, and mining-induced hazard assessment, but the extremely low signal-to-noise ratio (SNR) of field waveforms and the difficulty of preserving phase information continue to limit its practical value. Existing deep-learning denoisers primarily focus on amplitude fidelity while introducing measurable phase distortions that degrade downstream source-localization tasks. We present MRGAN, a multi-resolution residual generative adversarial network that integrates three-level Daubechies wavelet decomposition, multi-scale residual encoding with kernels of size 3, 5, and 7, a wavelet feature fusion module with channel attention, and a multi-scale PatchGAN discriminator. The network is optimized under a composite loss function combining Wasserstein adversarial loss with gradient penalty, L1 reconstruction loss, a novel Hilbert-transform-based phase-aware loss, and a wavelet-domain coefficient loss. Extensive experiments on three public benchmarks (STEAD, EGS Collab Experiment 2, and PNW-ML) show that MRGAN reaches 14.28 dB output SNR and 3.21 \(∘\) mean phase error at −5 dB input SNR—a 61.5% phase-error reduction over the DeepDenoiser baseline. On field data, MRGAN achieves a 10.72 dB SNR improvement while retaining 96.8% of microseismic events, and it generalizes to unseen event categories in PNW-ML without retraining. Inference runs in 12.85 ms per sample, making MRGAN directly deployable in operational monitoring pipelines.