Elements: A Convergent Physics-based and Data-driven Computing Platform for Building Modeling (09/2023 - present)
Team
Graduate Students
Introduction
Building modeling is used to establish computational models of their physical characteristics, indoor environments and energy use with external weather conditions acting as inputs. An accurate building model supports a variety of downstream applications, such as smart building management, retrofit analysis, and decarbonization. Current building modeling practice uses either physics-based and data-driven approaches. Physics-based methods model building dynamics using physical principles, which are sound and reliable, yet often have to compromise accuracy due to high computational cost, limited mechanistic rules, and incomplete input information. Data-driven approaches are computationally efficient and offer flexibility, but they lack interpretability and often are difficult to extrapolate, which hinders their field applications. This project proposes to overcome these gaps by developing a novel cyberinfrastructure with an advanced integration mechanism to fulfill convergent physics-based and data-driven modeling. Research outcomes intend to enable accurate and computationally efficient building modeling in practice, and thereby benefit many relevant applications whose goals are a sustainable and resilient built environment. Education and outreach activities, including interdisciplinary curriculum development, minority and K12 student engagement, are closely integrated into specific research activities.
This project will conduct unique, interdisciplinary research to design novel mechanisms that unify physics-based and data-driven modeling approaches. The project first plans to investigate and understand discrepancies between building models at different fidelities and actual measurements. Based upon developed understanding, several machine learning residual models and Neural Ordinary Differential Equations are developed to integrate physics-based and data-driven models. A flexible and easy-to-use cyberinfrastructure, including user interface layers and modeling layers, will be developed as an open-source platform to support convergent building modeling practice. The modeling framework and cyberinfrastructure are validated using field measurements from multiple resources. This project makes fundamental contributions to the building modeling field, as well as other engineering domains using diverse modeling approaches.
This award by the Office of Advanced Cyberinfrastructure is jointly supported by the Division of Chemical, Bioengineering, Environmental, and Transport Systems (ENG/CBET) and the Division of Civil, Mechanical & Manufacturing Innovation (ENG/CMMI).
This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.
link to NSF website
Community and Code Repositories
- Arbitrarily-Conditioned Multi-Functional Diffusion for Multi-Physics Emulation
- Efficient Kernel-Based Solvers for Nonlinear PDEs
- Invertible Fourier Neural Operators for Tackling Both Forward and Inverse Problems
- Standard-GP-is-all-you-need-for-HDBO
- Equation Discovery with Bayesian Spike-and-Slab Priors and Efficient Kernels
- Multi-Resolution Active Learning of Fourier Neural Operators
- Solving High Frequency and Multi-Scale PDEs with Gaussian Processes
- Functional Bayesian Tucker Decomposition for Continuous-indexed Tensor Data
Project Progress
- LLM Agent for Building Design: Large Language Models (LLMs) are gaining increasing attention for their potential to improve efficiency and sustainability in the building domain. Building Energy Modeling (BEM) traditionally requires substantial time and effort during the design and development phases. To alleviate this burden and promote the broader adoption of energy modeling in building design and operations, Automated Building Energy Modeling (ABEM) has emerged as a promising solution. We have developed an efficient fine-tuning framework to adapt LLMs for ABEM tasks. Leveraging Low-Rank Adaptation (LoRA) and a comprehensive training dataset containing 490,000 samples, our approach enables effective LLM customization while maintaining computational efficiency. Further improvements in training speed and resource usage are achieved through model quantization and mixed-precision training. Using this framework, we introduce EPlus-LLMv2, a platform capable of automatically generating complex building models with varying geometries, thermal zones, construction materials, hourly schedules, and operational parameters. Evaluations on 402 modeling cases demonstrate 100% modeling accuracy while reducing manual modeling effort by over 98%. To enhance usability, we also developed an interactive human-AI interface, making the platform more accessible to designers and engineers. Finally, we discuss key insights and future directions for customizing LLMs in ABEM and other applications within the building sector.
- AI Powered Physics-based Building Modeling: Establishing building energy models (BEMs) for design and analysis is fraught with challenges due to the intensive modeling efforts required, the expertise needed to use simulation software, and the necessary knowledge of building science. These factors make building modeling labor-intensive and hinder its widespread adoption in building development. To address these challenges and enhance automation in modeling practices, this work introduces Eplus-LLM as an automated building modeling platform. This platform leverages a fine-tuned large language mode to directly translate natural language descriptions of buildings into established models featuring various geometries, occupancy scenarios, and equipment loads. By fine-tuning the LLM, it is tailored to understand natural language and simulation requirements from users, converting human descriptions into EnergyPlus modeling files. The Eplus-LLM platform then automates the building modeling process by invoking the EnergyPlus engine's API to simulate the auto-generated model files and produce the desired simulation results.
- Bayesian spatial-temporal function decomposition: Building modeling often involves simulating and computing spatial-temporal functions, such as temperature and energy functions. In this work, we develop a Bayesian approach for learning and representing such functions. Specifically, we decompose the multi-variate spatial-temporal function as a combination of single-variable latent functions, where each latent function is over one spatial or temporal variable. To allow interpretability and easy learning, we combine those single-variable functions via a multi-linear Tucker decomposition form. We use Gaussian processes (GP) as functional priors to model the latent functions. Then, we convert each GP into a state-space prior by constructing an equivalent stochastic differential equation (SDE). This way, we can largely reduce the cost for training and inference. We use message massing and conditional moment match to develop an efficient inference algorithm for posterior approximation. The advantage of our method is shown in both synthetic real-world benchmark datasets. In synthetic data, our method can correctly recover the latent functions and reconstruct the response surface accurately within quite noisy observations. On the US temperature dataset that records interactions among latitude, longitude, and time, our learned latent functions not only reflect the spatial proximity of the locations in terms of both latitude and longitude, but also aligns with the climate change in the past 260 years.
- Bayesian differential equation discovery from temporal and spatial temporal systems: We aim to uncover interpretable patterns and summaries in real spatio-temporal building data. A common approach to characterizing these patterns is through differential equations, including ordinary differential equations (ODEs) and partial differential equations (PDEs). Consequently, we have developed a Bayesian governing equation discovery method called KBASS (Kernel-Based and Bayesian Spike-and-Slab priors). In our method, kernel regression is utilized to estimate the target function, providing a flexible, expressive, and robust solution to handle data sparsity and noise. This is combined with a Bayesian spike-and-slab prior, which is an optimal Bayesian sparse distribution, to facilitate effective operator selection and uncertainty quantification. For efficient posterior inference and function estimation, we have devised an expectation-propagation expectation-maximization (EP-EM) algorithm. To address the computational challenges of kernel regression, we place the function values on a mesh, inducing a Kronecker product construction, and leverage tensor algebra to enable efficient computation and optimization. Our method's effectiveness has been validated through various benchmark tasks, such as discovering the ODE from a chaotic system (Lorenz 96) and a complex fourth-order PDE (Kuramoto-Sivashinsky equation) from a spatio-temporal system. Additionally, our method has successfully identified a predator-prey equation that accurately characterizes the relationship between lynx and hares recorded by Hudson Bay Company from 1900 to 1920.
- Multi-resolution active neural operator learning: In building systems, quantities of interest (e.g., energy use) often evolve through complex processes that can be partially described by differential equations. However, solving high-fidelity equations is computationally expensive. This has led to the development of data-driven operator models for faster predictions. Among these, the Fourier Neural Operator (FNO) is a popular framework, known for its state-of-the-art performance and efficiency in training and prediction. However, gathering training data for FNOs is a costly bottleneck because it frequently requires expensive physical simulations. To address this issue, we propose the Multi-Resolution Active Learning of FNO (MRA-FNO). This method dynamically selects input functions and resolutions to minimize data costs while optimizing learning efficiency. Specifically, we introduce a probabilistic multi-resolution FNO and use ensemble Monte Carlo to develop an effective posterior inference algorithm. For active learning, we maximize a utility-cost ratio as the acquisition function, selecting new examples and resolutions at each step. To enable efficient utility computation, we use moment matching and the matrix determinant lemma. Additionally, we implement a cost annealing framework to prevent over-penalizing high-resolution queries early on. This over-penalization is a significant issue when cost differences between resolutions are large, often causing active learning to remain stuck at low-resolution queries and resulting in poor performance. Our method overcomes this challenge and is applicable to general multi-fidelity active learning and optimization problems. The advantages of our approach have been demonstrated in several benchmark operator learning tasks.
Related Publications
- Li, Shuai and Xu, Yifang and Chen, Jianli and Zhu, Siyao and Cai, Jiannan. (2025). A large language model-based platform for real-time building monitoring and occupant interaction. Journal of building engineering.
- Da Long, Zhitong Xu, Qiwei Yuan, Yin Yang, and Shandian Zhe, “ Invertible Fourier Neural Operators for Tackling Both Forward and Inverse Problems”, The 28th International Conference on Artificial Intelligence and Statistics (AISTATS), 2025
- Zhitong Xu, Haitao Wang, Jeff M. Phillips, and Shandian Zhe, “Standard Gaussian Process is All You Need for High-Dimensional Bayesian Optimization", International Conference on Learning Representations (ICLR), 2025
- Zhitong Xu, Da Long, Yiming Xu, Guang Yang, Shandian Zhe, and Houman Owhadi, “Toward Efficient Kernel-Based Solvers for Nonlinear PDEs”, Forty-Second International Conference on Machine Learning (ICML), 2025.
- Da Long, Zhitong Xu, Guang Yang, Akil Narayan, and Shandian Zhe, “Arbitrarily-Conditioned Multi-Functional Diffusion for Multi-Physics Emulation”, Forty-Second International Conference on Machine Learning (ICML), 2025.
- Jiang, G., Ma, Z., Zhang, L., & Chen, J. (2024). "EPlus-LLM: A large language model-based computing platform for automated building energy modeling", Applied Energy, 367, 123431.
- Shikai Fang, Madison Cooley, Da Long, Shibo Li, Robert M. Kirby, and Shandian Zhe, "Solving High Frequency and Multi-Scale PDEs with Gaussian Processes", Proceedings of The International Conference on Learning Representations(ICLR), 2024.
- Shibo Li, Xin Yu, Wei W. Xing, Robert M. Kirby, Akil Narayan, and Shandian Zhe, "Multi-Resolution Active Learning of Fourier Neural Operators", Proceedings of The 27th International Conference on Artificial Intelligence and Statistics (AISTATS), 2024
- Da Long, Wei W. Xing, Aditi S. Krishnapriyan, Robert M. Kirby, Shandian Zhe, and Michael W. Mahoney, "Equation Discovery with Bayesian Spike-and-Slab Priors and Efficient Kernels", Proceedings of The 27th International Conference on Artificial Intelligence and Statistics (AISTATS), 2024.
- Shikai Fang, Xin Yu, Zheng Wang, Shibo Li, Rboert M. Kirby, and Shandian Zhe, "Functional Bayesian Tucker Decomposition for Continuous-indexed Tensor", The International Conference on Learning Representations (ICLR), 2024.
- Xu, Y., Zhu, S., Chen, J., Cai, J., & Li, S. (2024). A GPT-Integrated Platform for Real-Time Building Monitoring and Occupant Interaction. Journal of Building Engineering, Under Review.