Deep Learning-Based Architectural Visual Style Transfer Simulation and Functional Performance Optimization

Abstract

Architectural style transfer methods have difficulty achieving a balanced optimization of both visual aesthetics and functional performance, resulting in frequently unpragmatic solutions that fail to meet real-life architectural needs. The paper explores the unique difficulties experienced by researchers engaged in architectural visual style transfer simulation and proposes a multi-objective optimization framework utilizing deep learning that addresses these difficulties. This paper addresses the issue of maintaining a high level of fidelity while adhering to physical constraints without jeopardizing structural safety. The proposed framework utilizes a CycleGAN (Cycle Generative Adversarial Network), which incorporates an attention mechanism
to increase the sensitivity to the transfer of local characteristics, and combines matrix representations with a VGG-19 (Visual Geometry Group 19) network to provide two types of loss (style and content) that ensure a high level of visual fidelity. The NSGA-II (Non-Dominated Sorting Genetic Algorithm II) algorithm is included in the framework to dynamically optimize both architectural style and functional performance by using energy use anddaylighting calculated by EnergyPlus and Radiance as hard constraints. Achieving this will require aclosed-loop generation-simulation-evaluation feedback process. Real-time imports of style transfer designs go into a Building Information Modeling (BIM) andVirtual Reality (VR) simulation where EnergyPlus and Radiance will compute the designs’ performance. The performance data are then used as feedback for the NSGA-II algorithm to
modify generator settings and produce an optimal design. Finally, the architectural model shows the trade-offs between the aesthetics, energy use, and daylight use of the various design solutions to provide designers with visual assistance for deciding upon which is best. Experimental results indicate that this approach resulted in better scores in user satisfaction surveys and 16.4% less EUI than pre-transfer models. The post-transfer sDA increased 8.2% as compared to StyleGAN and average use of space increased 22.3% over pre-transfer models. The average maximum deformation was 7.8 mm and average maximum stress was 18.1 MPa, with no detrimental effect to the safety of the building structure. The research provides both theoretical
and practical tools supporting intelligent building design that offers both aesthetic and practical value.

Keywords: Deep learning; style transfer; architectural simulation; architectural functional performance optimization; generative adversarial network

Cite As

J. Zhao, "Deep Learning-Based Architectural Visual Style Transfer Simulation and Functional Performance Optimization",
Engineering Intelligent Systems, vol. 34 no. 3, pp. 341-352, 2026.