ETD

• nuclear explosions
• seismic activity
• TDA

This thesis examines the contemporary challenge of distinguishing subterranean nuclear detonations from natural seismic events with contemporary machine-learning methodologies, focusing specifically on convolutional neural networks (CNN) and topological data analysis (TDA). Dependable seismic discrimination continues to be crucial in global nuclear test monitoring, particularly as event magnitudes diminish and waveform characteristics between explosions and earthquakes increasingly converge. While traditional physics-based discriminants, such as depth estimates, amplitude ratios, and waveform morphology, are still important, their consistent application in a variety of geological contexts and heterogeneous seismic networks often presents difficulties. Recent advancements in machine learning have the potential to enhance previous methods by directly learning discriminative patterns from seismic data. The study begins with a thorough review of the literature on machine learning applications in seismic discrimination, with a focus on convolutional neural networks as the most popular contemporary approach. CNN-based approaches have exhibited robust performance by acquiring hierarchical representations from waveform and spectral inputs, especially when trained on standardized datasets like GTUNE. This thesis examines research that highlights CNNs' benefits, like increased scalability and classification accuracy, as well as their disadvantages, like decreased interpretability and vulnerability to regional bias. The importance of incorporating physical context into data-driven models is highlighted by hybrid approaches that offer improved generalization by combining physics-based attributes with CNN-derived representations. The thesis underscores the need of data exploration and preparation as essential for effective machine learning. A variety of seismic datasets were analyzed, encompassing curated explosion-only datasets and mixed seismic catalogs. A comprehensive exploratory data analysis was performed to delineate waveform duration variability, temporal event distribution, geographic station coverage, and relationships among physically linked characteristics. Significant differences between earthquake and explosion signals were identified using time-domain waveform visualizations and frequency-domain power spectral density analyses, which informed the selection of feature representations for further modeling. This work primarily contributes by exploring Topological Data Analysis as a complementary and innovative approach for seismic discrimination. TDA offers mathematically rigorous tools for deriving geometric and structural insights from intricate data via persistent homology. This research illustrates that by converting seismic waveforms into geometric shapes and calculating persistence diagrams, topological summaries effectively encapsulate invariant structural characteristics that are resilient to noise, station effects, and signal distortion. These qualities mitigate the constraints of exclusively data-driven models and provide an interpretable framework for characterizing seismic waves. Machine-learning pipelines were established to evaluate traditional CNN-based classification against topological representations obtained from seismic data. CNN models were trained on preprocessed waveform and spectral inputs, demonstrating robust discriminating performance while underscoring problems associated with overfitting and validation. Concurrently, topological characteristics obtained from persistence diagrams have demonstrated their efficacy as meaningful representations for machine-learning categorization. The amalgamation of exploratory analysis, frequency-domain characterization, and topological feature extraction creates a robust framework for assessing seismic discrimination techniques. This thesis illustrates that the integration of machine learning with topological approaches offers a robust and adaptable strategy for seismic event discrimination. This research enhances the dependability, interpretability, and generality of nuclear explosion monitoring systems by anchoring data-driven models in physical understanding and geometric structure. The research team demonstrated that novel techniques, particularly Topological Data Analysis, can effectively enable machine learning to differentiate between seismic activity and underground nuclear explosions. The research team demonstrated that novel techniques, particularly Topological Data Analysis, can effectively enable machine learning to differentiate between seismic activity and underground nuclear explosions.