Synthetic Interpretations: AI-Driven Scoring Framework for Architectural Design Evaluation

Abstract

While artificial intelligence (AI) has significantly influenced architectural design through generative tasks like conceptual exploration and visualization, its capacity for nuanced qualitative evaluation remains underexplored. Effective evaluation requires a convergence of subjective interpretation and objective rigor, addressing contextual relationships, formal qualities, adherence to design principles, programmatic functions, construction strategies, structural systems, and sustainable practices. This research addresses these challenges by developing an AI-driven scoring framework, ArchiJury, based on synthetic architectural reviews aligned with established evaluation criteria. Two distinct AI models form the methodological basis of the study: The first employs visual transformer models and a synergy simulation algorithm for precise, context-based criterion-specific evaluation. The second uses a ResNet-18 deep-learning architecture for multi-criteria holistic scoring, trained end-to-end with an annotated dataset and optimized through mean squared error (MSE) loss, and utilizes Grad-CAM heatmaps for interpretability by visually representing the influential image regions guiding AI scoring decisions. The outputs of both models are comparatively discussed with human expert evaluations to critically assess AI’s potential and limitations, and implications of AI driven evaluation, clarifying how these computational methods align or diverge from expert judgment and exploring their significance for scalable, consistent, and nuanced architectural evaluation. © 2025, Education and research in Computer Aided Architectural Design in Europe. All rights reserved.