Shikang Wen / 温世康

Shikang Wen / 温世康

Research in building energy, computational design, and intelligent built environments.

AI-assisted generative building design and energy modeling

Published research Sep 2026

Overview

This project developed an AI-assisted workflow that translates high-level design inputs - a natural language description or a single building image - into EnergyPlus-compatible building geometry, connecting generative design with building energy simulation.

Four-stage workflow from a user prompt through an image generation model, a 2D image, an image-to-3D model, and EnergyPlus-compliant geometry conversion to an EnergyPlus model, with a feedback loop for further requirements.
The proposed workflow: from a text prompt or image to a simulation-ready EnergyPlus model, with iterative refinement.

Method

ChatGPT-image generates or preprocesses the 2D architectural representation, Hunyuan3D-2.1 reconstructs the 3D geometry, and a mesh-to-EnergyPlus method converts the reconstructed mesh into simulation-ready geometry. A graph neural network trained on a purpose-built synthetic dataset of 13,000 buildings classifies mesh faces into walls, windows, and roofs. The workflow was evaluated on multiple footprint shapes at three building scales across five Chinese climate zones.

Color-coded input images and the corresponding generated EnergyPlus models for low-rise, medium-rise, and high-rise buildings with rectangular, L-shape, U-shape, circular, courtyard, setback, modular, and curved forms.
Input images and the EnergyPlus models generated by the workflow across three building scales.

Key findings

Deterministic calibration kept gross floor area and window-to-wall ratio close to their target values, and energy use intensity (EUI) remained consistent with manually built reference models under identical boundary conditions. For a campus office building in Hong Kong reconstructed from a single image, the EUI differed from the reference model by about 3.3%. The tested cases were converted into EnergyPlus models in less than five minutes of machine-processing time.

A photograph of a campus office building is cleaned and color-coded with ChatGPT, reconstructed as a 3D mesh by Hunyuan3D-2.1, and converted into an EnergyPlus model that is compared with a manually reconstructed model.
Reconstructing a Hong Kong campus office building from a single image into an EnergyPlus model.

From one constrained text prompt, the workflow also generated six alternative designs for the same building, with EUI values 2.7%-3.1% above the reference, enabling rapid, energy-oriented comparison of conceptual design options.

Stacked bar chart of annual energy use intensity for the baseline design and six generated curved and modular designs, broken down into lighting and equipment, fans, pumps, and cooling.
Annual EUI of the original design and six alternatives generated by the workflow.

Contribution

Methodology, software, validation, investigation, formal analysis, data curation, and original-draft writing.