Exploring LLM-powered agents for modeling thermal dynamics of buildings

Modeling building thermal dynamics is essential for energy optimization, yet building heterogeneity and non-stationary dynamics demand per-building customization that typically requires expert intervention. Automated scientific discovery workflows powered by Large Language Models (LLM) could significantly decrease the human expertise requirements for generating custom thermal models at scale, but their performance varies considerably depending on configuration and task, raising a key question: how should we configure scientific agents for thermal dynamics modeling? To answer this, we developed ThermalForge, a platform for experimenting with agentic approaches to hybrid neural-physics modeling of thermal dynamics. Using a year-long dataset from hundreds of U.S. residential smart thermostats, we investigate the impact of physics knowledge in prompts, constraints on neural architecture search, LLM effectiveness in determining physics-neural transition points, model stochasticity, cost-accuracy tradeoffs, and performance against benchmark methods. Our results show that agentic modeling can match or exceed state-of-the-art automated methods while offering greater flexibility and reduced expertise requirements. We found important differences in configuring the physics-neural transition point, and demonstrate how the inherent variability of LLM outputs can be transformed from uncertainty into a mechanism for robust model discovery. This work presents the first large-scale application of LLM-powered modeling agents to real-world observational data, demonstrating their viability for thermal dynamics modeling and paving the way to scalable methods that embed domain expertise within automated workflows.