Artificial Intelligence as a Tool for Predicting the Aesthetic Appeal of Abstract Art: A Deep Learning Approach

This study proposes a novel deep learning framework for predicting the aesthetic appeal of abstract artworks, bridging the gap between subjective visual judgment and objective computational modeling. We introduce a Vision Transformer (ViT) architecture enhanced by contrastive learning, which jointly captures absolute and relative aesthetic cues. The model is trained on a hybrid dataset comprising over 5,000 abstract artworks, enriched with aesthetic scores derived from a fusion of crowdsourced evaluations, expert ratings, and predictive proxies. Optimization is performed using a Lookahead-Rectified Adam (RAdam) strategy to ensure fast convergence and generalization. Empirical results demonstrate that the proposed ViT model outperforms traditional CNN-based models such as ResNet, EfficientNet, and DenseNet by margins of up to 20% in <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">R</i><sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> score and 15% in MAE. Key performance metrics include an F1 score of 0.87, a Mean Absolute Error (MAE) of 10.0, and an <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">R</i><sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> score of 0.92. Additional analysis using scatter plots, Taylor diagrams, and violin plots confirms strong alignment with human judgment. The originality of this work lies in its mathematically grounded architecture for abstract aesthetic evaluation, its integration of contrastive loss for relational understanding, and its interpretability pipeline using attention-based visualization. The proposed model has applications in generative design, art recommendation, and creative AI interfaces.