Subsurface fracture network characterization and geothermal energy system design optimization

Abstract

This thesis addresses the critical challenges in the characterization, optimization, and design of fracture networks in enhanced geothermal systems (EGS) through the development and application of advanced machine learning and optimization techniques. The research focuses on improving the accuracy, efficiency, and scalability of fracture network characterization, geothermal heat extraction optimization, and multi-objective design of geothermal systems, contributing to the sustainable development of geothermal energy resources. The first part of the thesis introduces novel deep generative models for fracture network characterization. A variational auto-encoder and generative adversarial network (VAE-GAN) framework is proposed to estimate fracture networks by capturing the complex distribution of fracture parameters and incorporating prior geological knowledge. This framework maps high-dimensional fracture parameter distributions into low-dimensional continuous fields, significantly reducing uncertainty through Bayesian ensemble smoothing. Additionally, a pre-trained denoising diffusion model is developed for generative inversion of fracture networks, leveraging sparse and noisy observational data to improve reconstruction accuracy and provide a robust probabilistic framework for uncertainty quantification. These models demonstrate superior performance in numerical experiments and real-world applications, offering a transformative approach to subsurface fracture characterization. The second part of the thesis focuses on surrogate-assisted optimization algorithms for geothermal heat extraction and well-placement optimization. A Surrogate-assisted Level-based Learning Evolutionary Search (SLLES) algorithm is proposed to optimize well-control schemes, balancing exploration and exploitation to achieve superior heat sweep efficiency. Furthermore, a Generalized Data-driven Differential Evolutionary Algorithm (GDDE) is introduced to reduce the computational cost of well-placement and control optimization by leveraging probabilistic neural networks and radial basis function surrogates. These algorithms outperform state-of-the-art methods in benchmark functions and real-world geothermal systems, laying a solid foundation for efficient geothermal energy extraction. The third part of the thesis tackles high-dimensional, computationally expensive multi-objective optimization problems in geothermal system design. A Classifier-assisted Rank-based Learning and Local Model-based Evolutionary Algorithm (CLMEA) is proposed, utilizing uncertainty information and rank-based learning strategies to generate promising solutions and maintain diversity in high-dimensional decision spaces. Additionally, a Learning-based Multi-Objective Generative Model (LMOGM) is developed, replacing traditional genetic operators with an attention-enhanced convolutional residual network to generate high-quality candidate solutions. These algorithms demonstrate superior performance in benchmark problems and geothermal reservoir optimization, offering a scalable and efficient approach to multi-objective design. The final part of the thesis introduces a multi-fidelity machine learning framework with knowledge transfer to enhance the efficiency of geothermal energy system design and optimization. This framework leverages coarse and fine surrogate models to iteratively query informative data points, achieving significant speed-ups in optimization while maintaining accuracy. The proposed methods are validated through extensive experimental tests, demonstrating their capability to accelerate decision-making and improve the design of complex geothermal systems. Overall, this thesis advances the state-of-the-art in geothermal energy system design and optimization by integrating deep generative models, surrogate-assisted optimization algorithms, and multi-fidelity machine learning techniques. The proposed methodologies provide accurate, efficient, and scalable solutions for fracture network characterization, heat extraction optimization, and multi-objective design, paving the way for sustainable and efficient geothermal energy extraction. The contributions of this research have broad applicability to related fields, including fluid flow, reservoir management, and renewable energy systems, enabling accelerated discovery of optimal designs for complex systems.