Smith, David R.Liu, Qing HuoCui, Liangze2025-07-022025-07-022025https://hdl.handle.net/10161/32735<p>Electromagnetic (EM) inverse problems play a crucial role in various engineering and scientific domains, including geophysical exploration, EM devices design, biomedical imaging, and so on. Traditional methods for solving these problems rely heavily on computational electromagnetics (CEM) simulations and iterative optimization techniques, which are often computationally expensive and inefficient when dealing with high-dimensional design spaces. The growing complexity of EM problems necessitates novel approaches that can efficiently explore design spaces, extract meaningful patterns, and improve optimization accuracy. This dissertation investigates the integration of machine learning (ML) techniques into EM inverse problem-solving, aiming to enhance computational efficiency and improve design outcomes in two key application areas: EM inverse scattering problem (e.g. hydraulic fracturing detection) and EM inverse design problem (e.g. antenna design optimization).</p><p>To address the challenges associated with real-time hydraulic fracturing detection, we applied ML techniques to an EM-based monitoring system, developed as part of a U.S. Department of Energy-funded project. This system tracks dynamic changes in pressure and salinity within hydraulically fractured networks using EM field measurements. By leveraging ML models for data processing, noise reduction, and feature extraction, we significantly improved the system’s ability to interpret complex EM signals and reconstruct subsurface fracture conductivity distributions. However, discrepancies between EM simulations, field-collected data, and inversion results revealed the influence of streaming potential (SP) effects. To mitigate these discrepancies, we should require a multi-physics modeling approach that accounts for both EM and SP effects in future research, improving the accuracy of inversion results and enhancing hydraulic fracture characterization.</p><p>In addition to subsurface monitoring, this dissertation explores ML-driven methods for antenna design optimization, where traditional optimization techniques suffer from data inefficiency and high computational costs. We developed an active learning-based framework that combines K-Nearest Neighbors (KNN) with adaptive sampling strategies to reduce dataset requirements while maintaining high optimization accuracy. Unlike conventional ML approaches that rely on pre-constructed datasets, our method dynamically updates training data in real-time through iterative interaction with full-wave EM solvers such as High Frequency Structure Simulator (HFSS). This approach resulted in up to a 30-fold improvement in computational efficiency compared to traditional methods such as artificial neural networks (ANNs) and Bayesian optimization.</p><p>Furthermore, we introduce a generative AI-based method for complex antenna design using Auxiliary Classifier Generative Adversarial Networks (AC-GANs). The AC-GAN framework consists of a discriminator-generator model, where the discriminator predicts class labels from geometric models, and the generator creates novel antenna structures that satisfy predefined performance criteria. We incorporate a filtering layer to refine design candidates and employ active learning-inspired evolutionary strategies to improve data efficiency. The proposed approach successfully optimized antennas (e.g. dipole antennas, filtering antennas), with experimental validation confirming the feasibility and effectiveness of the generated designs.</p><p>The findings of this research demonstrate that ML can effectively tackle EM inverse problems by accelerating computational workflows, reducing simulation burdens, and enabling data-driven design exploration. By integrating ML into hydraulic fracturing detection and antenna optimization, this dissertation advances the intersection of AI and CEM, paving the way for more efficient, scalable, and practical solutions to complex engineering challenges.</p>https://creativecommons.org/licenses/by-nc-nd/4.0/Electrical engineeringElectromagneticsAntennaElectromagneticsFracturingInverse problemsMachine learningDissertation