Parallel typological optimization in architectural design: a ResNet-enhanced island-based evolutionary algorithm

Evolutionary optimization has become a widely adopted approach in performance-based building design. It enables the discovery of high-fitness solutions through iterative search processes. However, its application still faces significant challenges in early-stage architectural design. These challenges arise from difficulties in systematically exploring diverse and competing conceptual directions. To address these limitations, this study introduces MGA_R, a hybrid evolutionary algorithm designed for multi-typological parallel optimization in building design. The proposed algorithm integrates a ResNet-based building typology classification model with an island-based coevolution model. In addition, MGA_R is incorporated with the EvoMass framework to enable the co-evolution of various building typologies within a single process. By leveraging the predictive capabilities of the classification model to categorize architectural forms and combining it with an approach that manages multiple subpopulations across distinct typologies, MGA_R facilitates more efficient exploration while maintaining diversity in design directions. Two case studies are presented to evaluate both the search efficiency of MGA_R compared to other optimization methods and its effectiveness in supporting iterative design development. The results demonstrate how machine learning and computer vision techniques enhance evolutionary algorithms for architectural performance-based design, highlighting their synergistic potential in bridging conceptual ideation with technical refinement during early design phases.