01 Explainable AI-driven framework for automated design innovation assessment: A hybrid deep learning approach for creative work evaluation

Abstract

Design innovation assessment represents a critical challenge in contemporary creative industries, where traditional evaluation methods rely heavily on subjective expert  judgment, leading to inconsistencies, inefficiencies, and limited  scalability.  The emergence of  artificial  intelligence   technologies   offers   unprecedented   opportunities to revolutionize design evaluation processes, yet existing approaches  suffer  from black-box limitations that hinder adoption in professional design contexts where transparency and interpretability are paramount. Here we present an  explainable  AI- driven framework that synergizes advanced deep  learning  architectures  with interpretable machine learning techniques to enable automated, objective, and transparent design  innovation  assessment.  Our  hybrid  approach  integrates DenseNet201 for comprehensive visual feature extraction with Support Vector Machine classification for  robust  decision  boundary  formation,  enhanced  by  multiple explainable AI techniques including Gradient-weighted Class Activation Mapping, Integrated Gradients, and Layer-wise Relevance Propagation to provide multi-level interpretability. Through comprehensive evaluation on a curated dataset of 5,247 design works spanning product design, graphic design, architectural design, and user interface design,  our  framework  achieves  exceptional  performance  with  97.8% accuracy,  96.4% precision,  97.1%  recall,  and  96.7%  F1-score.   The   explainability analysis demonstrates that Layer-wise Relevance  Propagation  provides  the  most effective interpretability  with  95.6%  innovation  element  localization  accuracy  and  93.4% expert acceptance rate. User studies involving 30 design experts and  120 professional designers confirm significant improvements in evaluation  efficiency  (42% time reduction) and consistency (92% inter-rater agreement vs. 67% for traditional methods). This framework establishes a new paradigm for design evaluation  that  combines computational precision with  human-interpretable  insights,  offering substantial potential  for  transforming  design  education,  creative  industry  workflows, and innovation management practices.