Yingjie Liu Archives - Research Portal

An AI-augmented urban building energy Modeling framework integrating physics-based simulation, machine learning, and computer vision: The CityEnerTwin UBEM module

Abbasabadi, N., Worthy, A., Zhang, Z., Zhao, Y., & Liu, Y. (2026). An AI-augmented urban building energy Modeling framework integrating physics-based simulation, machine learning, and computer vision: The CityEnerTwin UBEM module. Energy and Buildings, 369, Article 117984. https://doi.org/10.1016/j.enbuild.2026.117984

View Publication

Abstract

Urban Building Energy Modeling (UBEM) plays a critical role in evaluating urban decarbonization strategies, but its accuracy and scalability are often constrained by simplified building archetype, limited façade-level data, and the computational cost of large-scale scenario analysis. This study presents the CityEnerTwin UBEM module, an artificial intelligence (AI)-augmented, physics-based framework that integrates modular workflows for computer vision–based façade semantic enrichment, localized archetype development, predictive and surrogate machine learning (ML), and climate-informed analysis to support scalable and computationally efficient urban energy modeling. Specifically, these workflows combine validated Vision Transformer (ViT)-based façade feature extraction methods for deriving material properties and window-to-wall ratio (WWR) from street-view imagery, K-means clustering for localized archetype development, and two complementary ML modeling streams: predictive models trained on measured operational energy data and surrogate models trained on synthetic UBEM simulation data for rapid retrofit scenario evaluation. Demonstrated on the University of Washington campus, the framework reduced campus-scale simulation error relative to measured energy consumption from 27.8% to 15.8%, while image-derived WWR inputs improved agreement for representative buildings from 23% to 20%. Predictive modeling using Random Forest (RF) achieved a test R2 of 0.88 and identified occupancy, building scale, and geometric efficiency as the dominant predictors of energy demand. Surrogate modeling using an Artificial Neural Network (ANN) achieved a test R2 of 0.98 and identified roof insulation, WWR, and shading depth as the most influential retrofit parameters, enabling rapid retrofit scenario evaluation. The results demonstrate the potential of integrated physics-based and data-driven workflows to improve UBEM representation, scalability, and computational efficiency for urban decarbonization planning. As the analytical core of the broader CityEnerTwin platform, the module connects spatially indexed building data and model outputs with interactive visualization to support urban-scale energy analysis and decision making.

Keywords

Urban building energy Modeling; Physics-based simulation; Computer vision; Machine learning; Surrogate Modeling; Retrofit scenario analysis

Enhancing urban building energy models with Vision Transformers: A Case study in material classification from Google street view

Liu, Y., & Abbasabadi, N. (2025). Enhancing urban building energy models with Vision Transformers: A Case study in material classification from Google street view. Energy and Buildings, 333, Article 115457. https://doi.org/10.1016/j.enbuild.2025.115457.

View Publication

Abstract

The growing urbanization and increased urban energy consumption highlight the need for energy use and greenhouse gas emissions reduction strategies. Urban Building Energy Modeling (UBEM) emerged as a valuable tool for managing and optimizing energy consumption at the neighborhood and city scales to support carbon reduction goals. However, the accuracy of the UBEM is often limited by the lack of large-scale building façade material dataset. This study introduces a new approach to enhance UBEM by integrating an automatic deep learning material classification pipeline. The pipeline leverages multiple views of Google Street View Images (SVIs) to extract building façade material information, utilizing two Swin Vision Transformer (ViT) models to capture both global and local features from the SVIs. The pipeline achieved a main material classification accuracy reached 97.08%, and the sub-category accuracy reached 91.56% in a multi-class classification task. As the first study to apply a deep learning model for material classification to enhance the UBEM framework, this work was tested on the University of Washington campus, which features diverse facade materials. The model demonstrated its effectiveness by achieving an overall accuracy increase of 11.4% in year-round total operational energy simulations. The scalability of this material classification pipeline enables a more accurate and cost-effective application of UBEM at broader urban scales.

Yingjie Liu

My interest lies in urban-scale building energy modeling and carbon accounting for climate mitigation. Specifically, I am focused on how digital documentation of the built environment can automate and enhance the accuracy of current accounting methods. Moreover, I am intrigued by how these advancements enable the broader application of bottom-up accounting approaches, informing early-stage design and influencing energy policy decisions.