Embedding Space Optimization For Enhanced Seismic Monitoring SEP-195 (2025)

Downloads

Min Jun Park Thesis

Table of contents

• Chapter 1. Introduction
  • 1.1 Seismic Monitoring Overview
  • 1.2 Deep Learning Foundations
  • 1.3 Thesis Contributions
  • 1.4 Thesis Structure
• Chapter 2. Embedding Space Optimization Theory
  • 2.1 Convolutional Neural Networks (CNNs)
  • 2.2 Embedding Techniques
  • 2.3 Optimization Strategies for Embeddings
• Chapter 3. Geological Carbon Storage Monitoring
  • 3.1 Background and Importance
  • 3.2 Methods and Metrics
  • 3.3 Synthetic Data Example
  • 3.4 Real-World Case Study: Aquistore
  • 3.5 Discussion and Insights
• Chapter 4. DAS-Based Seismic Event Detection
  • 4.1 Introduction to DAS Technology
  • 4.2 Challenges with Imbalanced Data
  • 4.3 Proposed Embedding Optimization Method
  • 4.4 Case Study: Microseismic Events
  • 4.5 Case Study: Earthquake Detection
• Chapter 5. Conclusions

Abstract

Seismic monitoring is essential for understanding subsurface processes and ensuring the safety and efficiency of geological applications such as carbon sequestration and seismic event detection. However, it faces challenges stemming from the complexity and large volume of seismic data. This dissertation addresses these challenges by utilizing embedding space optimization, a machine learning technique that transforms high-dimensional data into a lower-dimensional space, enhancing feature representation for enhanced seismic monitoring.

The first challenge involves detecting subsurface changes due to carbon dioxide (CO2) injection in geological carbon storage (GCS) projects. Obtaining accurate labels or ground truth information about the exact distribution of injected CO2 in the subsurface is impractical. To overcome this, we propose an unsupervised anomaly detection approach using embedding space optimization. By learning the normal patterns of the subsurface from pre-injection seismic data, the model identifies deviations in post-injection data as anomalies, effectively detecting CO2 plume signals without requiring labeled data. This method enhances the monitoring of GCS projects by providing a practical solution for identifying subsurface changes associated with CO2 injection.

The second challenge pertains to the detection of seismic events within datasets where seismic events are rare and noise is abundant. This rarity of seismic events poses difficulties for traditional supervised learning methods, which often fail to detect infrequent but critical seismic events. To address this, we introduce a semi-supervised clustering framework utilizing embedding space optimization. By combining a small amount of labeled data with a large volume of unlabeled data, the model effectively distinguishes between seismic events and noise. This approach enhances the detection performance of rare seismic events without the need for additional labeled data. We demonstrate the effectiveness of the proposed methods through two case studies in distributed acoustic sensing (DAS) data analysis: microseismic monitoring during hydraulic fracturing operations and earthquake detection in urban environments. In both cases, embedding space optimization significantly improves the models' ability to detect seismic events despite the scarcity of event data and without requiring additional training data.

The contributions of this dissertation are twofold. First, it advances the field of GCS monitoring by providing an unsupervised anomaly detection method that effectively identifies CO2 plumes without labeled data. Second, it enhances seismic event detection in datasets with rare events by using a semi-supervised clustering framework, improving detection accuracy of rare events in large volumes of noise-dominated data without the need for extensive labeled datasets. These methodologies leverage embedding space optimization to extract meaningful representations from complex seismic data, facilitating better interpretation and decision-making in geophysical applications.

Overall, this work demonstrates the potential of embedding space optimization to address critical challenges in seismic monitoring. By improving data analysis and detection capabilities without the need for additional labeled data, the findings contribute to the advancement of geophysical research and offer practical solutions for real-world applications in seismic monitoring.

Reproducibility and source codes

This thesis has been verified for reproducibility. The source code and detailed instructions for reproducing the results are available in the following GitHub repositories for each case study:

Case1 (chapter3): https://github.com/minjun1/DeepNRMS_Synthetic
Case2 (chapter4): https://github.com/minjun1/ESO_SDA1

Defense

Min Jun Park Defense PPT