Earthquake magnitude and location estimation from real time seismic waveforms with a transformer network

Abstract

Summary Precise real time estimates of earthquake magnitude and location are essential for early warning and rapid response. While recently multiple deep learning approaches for fast assessment of earthquakes have been proposed, they usually rely on either seismic records from a single station or from a fixed set of seismic stations. Here we introduce a new model for real-time magnitude and location estimation using the attention based transformer networks. Our approach incorporates waveforms from a dynamically varying set of stations and outperforms deep learning baselines in both magnitude and location estimation performance. Furthermore, it outperforms a classical magnitude estimation algorithm considerably and shows promising performance in comparison to a classical localization algorithm. Our model is applicable to real-time prediction and provides realistic uncertainty estimates based on probabilistic inference. In this work, we furthermore conduct a comprehensive study of the requirements on training data, the training procedures and the typical failure modes. Using three diverse and large scale data sets, we conduct targeted experiments and a qualitative error analysis. Our analysis gives several key insights. First, we can precisely pinpoint the effect of large training data; for example, a four times larger training set reduces average errors for both magnitude and location prediction by more than half, and reduces the required time for real time assessment by a factor of four. Second, the basic model systematically underestimates large magnitude events. This issue can be mitigated, and in some cases completely resolved, by incorporating events from other regions into the training through transfer learning. Third, location estimation is highly precise in areas with sufficient training data, but is strongly degraded for events outside the training distribution, sometimes producing massive outliers. Our analysis suggests that these characteristics are not only present for our model, but for most deep learning models for fast assessment published so far. They result from the black box modeling and their mitigation will likely require imposing physics derived constraints on the neural network. These characteristics need to be taken into consideration for practical applications.

Transfer learning: improving neural network based prediction of earthquake ground shaking for an area with insufficient training data

“…Li et al, 2018) or rapid earthquake characterization (e.g. Böse et al 2012, Hsu et al 2016, Ochoa et al 2018, Saad et al 2020, van den Ende and Ampuero 2020, Münchmeyer et al 2021, Zhang et al 2021. Böse et al (2008) have approached the problem of rapid earthquake characterization for EEWS using multi-station waveforms to extract a series of chosen parameters and use them as inputs for a feedforward neural network.…”

A probabilistic view on rupture predictability: all earthquakes evolve similarly

“…To resolve this issue, we apply our approach to teleseismic P arrival waveforms, which in contrast to STFs contain full spectral information up to ∼ 1 Hz. As neural network we adapted TEAM-LM (Münchmeyer et al, 2021) and apply it to a catalog of ∼35,000 events with nearly 750,000 manually labeled first P arrivals (Figure S8).…”

“…We collated a dataset of ∼ 35, 000 event with ∼ 750, 000 manually labelled first P arrivals. As neural network we adapted TEAM-LM (Münchmeyer et al, 2021). TEAM-LM consists of a combination of convolutional layers, a transformer network, and a mixture density output and predicts the event magnitude directly from the seismic waveforms at a flexible set of input stations.…”

“…Compared to the STF model, predictions P(M |O t ) and P(M |O M ) show higher uncertainties and systematic underestimation of the largest magnitudes at all times (Figure 4g). Higher uncertainties result from the fact that assessing magnitude from waveforms is harder than from STFs, whereas underestimation can be attributed to data sparsity (Münchmeyer et al, 2021). Due to the higher model uncertainties, all tails look rather like exponentially modified Gaussians than exponential distributions, as for the STF case.…”

“…We train for 100 epochs and use from each ensemble member the model with lowest validation loss for evaluation. Münchmeyer et al (2021) showed that neural network models suffer from data sparsity for large events, causing systematic underestimation of magnitudes. As a simple mitigation, they suggested upsampling these events in training, i.e., artificially increasing their occurrence.…”