An immediate report of the source focal mechanism with full automation after a destructive earthquake is crucial for timely characterizing the faulting geometry, evaluating the stress perturbation, and assessing the aftershock patterns. Advanced technologies such as Artificial Intelligence (AI) has been introduced to solve various problems in real-time seismology, but the real-time source focal mechanism is still a challenge. Here we propose a novel deep learning method namely Focal Mechanism Network (FMNet) to address this problem. The FMNet trained with 787,320 synthetic samples successfully estimates the focal mechanisms of four 2019 Ridgecrest earthquakes with magnitude larger than Mw 5.4. The network learns the global waveform characteristics from theoretical data, thereby allowing the extensive applications of the proposed method to regions of potential seismic hazards with or without historical earthquake data. After receiving data, the network takes less than two hundred milliseconds for predicting the source focal mechanism reliably on a single CPU.

The stability and robustness of determining earthquake magnitude are of great significance in earthquake monitoring and seismic hazard assessment. The routine workflow of determining earthquake local magnitude, such as the widely used Richter magnitude, may result in an unreliable measurement of earthquake magnitude because it relies on individual amplitude measurement of a single station, which is prone to be influenced by natural impulsive noise or anthropogenic noise. In this study, we present an automated estimation of earthquake magnitude by applying a deep‐learning algorithm named magnitude neural network (MagNet) based on the full‐waveform recordings from a network of seismic stations at China seismic experimental site (CSES). The MagNet consists of a compression component that extracts the global features of waveform data and an expansion component that yields a Gaussian probability distribution representing the magnitude estimation. The MagNet is trained with an augmented data set, which includes 21,700 training samples with evenly distributed magnitudes. From the prediction results on the test data set, the mean errors and standard deviations are −0.017 and 0.21, respectively, for 600 moderate earthquakes with magnitudes ranging from 3 to 5.9, and −0.011 and 0.14, respectively, for 70 small earthquakes with magnitudes ranging from 2.3 to 3.5. However, it remains challenging for large earthquakes (⁠ $magnitude>6.5$⁠), due to the lack of sufficient historical large earthquakes as training data. In addition, testing results show that the new method is capable of minimizing the impact of abnormal noises in the data. These results demonstrate the validity and merits of the proposed deep‐learning method in predicting earthquake magnitude automatically.

Recent advances in artificial intelligence allow seismologists to upgrade the workflow for locating earthquakes. The standard workflow concatenates a sequence of data processing modules, including event detection, phase picking, association, and event location, with elaborately fine‐tuned parameters, lacking automation and convenience. Here, we leverage deep reinforcement learning and develop a state‐of‐the‐art earthquake robot (EQBot) to help advance automated earthquake location. The EQBot learns from tremendous trial‐and‐error explorations, which aims to best align the observed P and S waves, complying with the geophysical principle of gather alignments in source imaging. After training on earthquakes (M ≥ 2.0) for a decade in the Los Angeles region, it can locate earthquakes directly from waveforms with mean absolute errors of 1.32 km, 1.35 km, and 1.96 km in latitude, longitude, and depth, respectively, closely comparable to the cataloged locations. Moreover, it can automatically implement quality control by examining the alignments of P and S waves. Our study provides a new solution to advance the earthquake location process toward full automation.

Migration-based earthquake location methods may encounter the polarity reversal issue due to the non-explosive components of seismic sources, leading to an unfocused migration image. Such a problem usually makes it difficult to accurately retrieve the optimal location from the migrated source image. In this study, by taking advantage of the general pattern recognition ability of the convolutional neural network, we propose a novel deep-learning image condition (DLIC) to address this issue. The proposed DLIC measures the goodness of waveform alignments for both P and S waves, and it follows the geophysical principle of seismic imaging that the best-aligned waveforms represent fully a best-imaged source location. A synthetic test shows that the DLIC can effectively overcome the polarity reversal issues. Real data applications to southern California show that the DLIC can enhance the focusing of the migrated source image over the classic source scanning algorithm. Further tests show that the DLIC applies to continuous seismic data, to regions with few previously recorded earthquakes, and has the potential to locate small earthquakes. The proposed DLIC shall benefit the migration-based source location methods.

Earthquake focal mechanisms are important for characterizing the subsurface faulting geometry and evaluating stress distributions. Existing approaches usually strive to determine the absolute focal mechanisms and may be subject to large uncertainties due to incomprehensive knowledge of the velocity model, particularly for moderate‐to‐small earthquakes. Alternatively, difficulties that arise from the velocity model can be largely mitigated by inverting the relative data variations in a series of earthquakes, because effects from the velocity model are systematic among all events in the vicinity. In this study, we propose a novel relative focal mechanism inversion (RFMI) method to invert the second‐order variations in a series of focal mechanisms utilizing a well‐constrained primary event. We test the RFMI method on both synthetic data and 251 real earthquakes (M ≥3) in the 2019 Ridgecrest sequence. The synthetic test results show that the RFMI method is robust and insusceptible to location errors (<2 km) and systematic velocity errors (5%). The real data application results demonstrate improved consistency among the inverted focal mechanisms, resulting in better characterization of the fault orientations than the Southern California Seismic Network (SCSN) focal mechanism catalog. The retrieved earthquake depths are also well correlated with the depths of the $Mw$ 6.4 and 7.1 mainshocks. Waveform cross‐correlation analysis verifies the reliability of the results. Furthermore, dynamic stress monitoring is enabled with decent resolution. The proposed RFMI method paves a new path toward achieving a rich number of reliable earthquake focal mechanisms, which will benefit the investigation of the earthquake process.

Earthquake data are one of the key means by which to explore our planet. At a large scale, the layered structure of the Earth is revealed by the seismic waves of natural earthquakes that go deep into its inner core. At a local scale, seismology for exploration has successfully been employed to discover massive fossil energies. As the volume of recorded seismic data becomes greater, intelligent methods for processing such a volume of data are eagerly anticipated. In particular, earthquake focal mechanisms are important for assessing the severity of tsunamis, characterizing seismogenic faults, and investigating the stress perturbations that follow a major earthquake. Here, we report a novel deep reinforcement learning method for inverting the earthquake focal mechanism. Unlike more typical deep learning applications, which require a large training dataset, a deep reinforcement learning system learns by itself. We demonstrate the validity and efficacy of the proposed deep reinforcement learning method by applying it to the Mw 7.1 mainshock of the Ridgecrest earthquakes in southern California. In the foreseeable future, deep learning technologies may greatly contribute to our understanding of the oceanographic process. The proposed method may help us understand the mechanism of marine earthquakes.