This doctoral research aims to develop and validate a machine-learning approach for detecting software antipatterns during the construction phase. The study addresses three central challenges: ambiguous taxonomies and definitions of antipatterns, elevated false-positive rates in detection, and limited reproducibility of learning-based results without standardized protocols. The proposed pipeline consolidates taxonomy and formal definitions, operationalizes antipatterns as logical compositions of code smells and measurable structural properties, and derives observable indicators from structural metrics. Supervised learning with explicit hyperparameter optimization and a metric–label correlation layer are planned to support robustness, interpretability, and traceable evidence.