Sequential pattern recognition in CAD operations: A deep learning framework for next-action prediction

Predicting the next action in Computer-Aided Design (CAD) workflows is a challenging task due to the complex and sequential nature of design operations. This study proposes a deep learning framework for sequential pattern recognition in CAD command prediction by integrating convolutional sequence modeling, semantic embeddings, and engineered distance-based features. Using the SketchGraphs dataset, CAD operations are modeled as sequences of parametric design commands and transformed into input–target pairs for next-action prediction. The proposed approach is evaluated against traditional methods, including frequency-based and Markov models, as well as deep learning architectures such as Caser and a Tiny-Transformer. Experimental results demonstrate that deep learning models significantly outperform traditional approaches, with the CNN-based Caser model achieving the best performance. In particular, the integration of Multilingual Universal Sentence Encoder (MUSE) embeddings with distance-based features yields the highest accuracy of 59.02%, surpassing the strongest traditional baseline by 7.6% and outperforming the Transformer-based model by 2.5%. These findings highlight the effectiveness of convolutional architectures in capturing structured sequential patterns in CAD operations and challenge the assumption that attention-based models universally dominate sequence modeling tasks. The proposed framework provides a robust foundation for intelligent CAD systems and next-action recommendation.