Leveraging synthetic data for deep learning denoising and prediction of measured earthquake waveforms

Alexander Bauer , Conny Hammer

Single-station recordings of teleseismic earthquakes are inherently complex due to the superposition of numerous seismic phases and their contamination with noise, which can be particularly problematic in urban environments. A detailed knowledge of the wavefield generated by teleseismic earthquakes is critical for high-precision research facilities like those involved in photon science or gravity wave detection. However, seismological stations are often sparsely distributed, especially in regions with low seismic hazard, such as Northern Germany. To address this, we introduce a deep learning approach to predict low-frequency earthquake waveforms at arbitrary locations within and around the Hamburg metropolitan area. Our aim is to train a convolutional neural network (CNN) to predict measured earthquake waveforms from synthetic ones. For this, we use measured earthquake waveforms from a seismic station close to Hamburg that has been in operation for almost 30 years. However, using the measured earthquake waveforms as labels requires denoising them first. Hence, we propose a two-step strategy: first, we generate noise-free synthetic waveforms, add artificial noise that emulates the station's actual noise characteristics, and train a first CNN to denoise them. Second, we apply the first CNN to the noisy measured waveforms to obtain noise-free labels. With these, we train a second CNN to translate synthetic waveforms into the noise-free measured ones. Applications of the second CNN to earthquakes not included in the training dataset show that this method effectively predicts measured waveforms not only for the training station but also for stations in and around Hamburg. This approach represents a significant step towards accurately modeling the seismic wavefield in three dimensions without the need for densely distributed seismological stations.