de Hoop, Maarten2023-01-032023-06-012022-122022-12-02December 2Jasperson, Hope. "Physics-based machine learning for classifying, forecasting, and blindly locating seismic events." (2022) Diss., Rice University. <a href="https://hdl.handle.net/1911/114200">https://hdl.handle.net/1911/114200</a>.https://hdl.handle.net/1911/114200The use of machine learning (ML) in seismology has skyrocketed within the past several years. However, studies often use generic networks that do not incorporate the physics of the specific task. Failure to use existing knowledge makes these problems more difficult and artificially reduces accuracy. Here, we focus on several problems from the seismic data processing pipeline and build networks that are specific to each problem and are grounded in physics. First, we present a method using unsupervised classification and an attention network to forecast labquakes using acoustic emission (AE) waveform features. We analyzed the temporal evolution of AEs generated throughout several hundred laboratory earthquake cycles and used a Conscience Self-Organizing Map (CSOM) to perform topologically ordered vector quantization based on waveform properties. The resulting map was used to interactively cluster AEs. We examined the clusters over time to identify those with predictive ability. Finally, we used a variety of LSTM and attention-based networks to test the predictive power of the AE clusters. By tracking cumulative waveform features over the seismic cycle, the network is able to forecast the time-to-failure (TTF) of lab earthquakes. Our results show that analyzing the data to isolate predictive signals and using a more sophisticated network architecture are key to robustly forecasting labquakes. Next, we tackled phase association by framing it as a combinatorial optimization problem and using reinforcement learning (RL). By leveraging the latent structure of typical problem instances, RL works even for problems that are computationally hard in the worst case. RL usually relies on domain-specific heuristics crafted over years of research. For many relevant problems, however, finely-tuned heuristics do not exist. We propose Generalized Optimizer for Unsupervised Deep Assignment (GOUDA), an unsupervised framework for seismic phase association. In place of a heuristic, GOUDA uses a deep consensus network to learn the latent structure of the data - the wave speed model - which is used to calculate rewards. We show on synthetic data of varying complexity that GOUDA effectively solves the association problem while simultaneously recovering an accurate 3D wave speed model and earthquake locations. Since this is achieved with minimal a priori information, GOUDA is the first-of-a-kind tool suitable for ab initio seismic imaging and phase association.application/pdfengCopyright is held by the author, unless otherwise indicated. Permission to reuse, publish, or reproduce the work beyond the bounds of fair use or other exemptions to copyright law must be obtained from the copyright holder.seismologycombinatorial optimizationreinforcement learningmachine learningdeep learningearthquake predictionearthquake forecastingphase associationassignment problemacoustic emissionThesis2023-01-03