Date of Award

6-26-2026

Date Published

July 2026

Degree Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Mechanical and Aerospace Engineering

Advisor(s)

Bing Dong

Keywords

Agentic AI;Building Control Optimization;Building Energy Modeling;Building to Grid Integration;Physics-Informed Machine Learning

Subject Categories

Engineering | Mechanical Engineering

Abstract

Buildings account for a significant share of global energy consumption, and meeting the 2050 net-zero decarbonization targets requires retrofitting approximately 10,000 buildings per day in the United States alone. However, current human-centered workflows for building modeling, simulation, control, and operation remain too slow and labor-intensive to support deployment at this scale. This dissertation proposes an integrated two-pillar framework consisting of a physics-informed machine learning framework for integrated building modeling, control, and simulation, and a multi-agent agentic AI framework for automating building energy engineering workflows built on top of that foundation. The first pillar introduces a Physics-Informed Modularized Neural Network (PI-ModNN) framework that integrates physical laws with data-driven learning for scalable building energy system modeling. The framework adopts a hierarchical modular architecture, progressing from envelope-level heat transfer modules to zone-level thermal dynamic models, multi-zone graph-based representations, and integrated building–HVAC–distributed energy resource systems through the BESTOpt simulation platform. The modular structure supports building retrofit analysis with physical consistency constraints, enabling reliable out-of-distribution predictions that unconstrained data-driven models fail to provide. The framework scales from single-zone to 38-zone buildings and further to a 30-building residential cluster, while preserving component-level fidelity in integrated building-HVAC-DER simulation. Based on this physics-informed digital twin foundation, this dissertation further demonstrates its value for advanced building control applications. Differentiable predictive control developed within the PI-ModNN framework achieves 21.1% energy savings, 42.5% energy cost reduction, and 52.2% peak load reduction. In addition, a deep reinforcement learning controller trained in the physics-informed environment and deployed in a real building achieves 31.4% coil-side HVAC load reduction and 28.4% peak load shifting. The second pillar develops a multi-agent agentic AI framework to automate building energy engineering workflows. The framework coordinates specialized large language model agents for tasks such as data processing, model configuration, training, control deployment, and performance evaluation through integration with BESTOpt via the Model Context Protocol. A large-scale benchmark comprising 6,156 runs is further conducted to systematically investigate how orchestration structure, model-size configuration, task complexity, and orchestrator–specialist coordination affect workflow accuracy, efficiency, and reliability. The results show that the two-stage centralized orchestration mode achieves 93–95% task accuracy, demonstrating that structured planning and orchestration design are critical to the performance of engineering-oriented agentic AI systems. To demonstrate the integrated capabilities of the proposed framework, two representative case studies are conducted. The first focuses on cluster-level modeling, simulation, and analysis in a 20-building residential community, demonstrating how the agentic AI framework can autonomously coordinate multi-step workflows to assess the impacts of retrofit and control strategies on energy use, operating cost, thermal comfort, and flexibility. The second focuses on optimal control development, where the BESTOpt environment is used to benchmark deep reinforcement learning algorithms driven by proposed agentic AI framework for coordinated building-DER-grid operation. Among the evaluated methods, the Soft Actor-Critic algorithm achieves the most balanced performance, reducing operating cost by 24.8%, grid import by 33.0%, peak demand by 41.9%, and PV curtailment by 87.3% compared with a rule-based baseline. Together, these two pillars establish an integrated foundation for a paradigm shift from conventional human-centered building engineering workflows toward agentic AI systems integrated with physics-informed digital twins, providing a scalable pathway toward future autonomous buildings and energy systems.

Access

Open Access

Download

Share

COinS