MACHINE LEARNING FOR COMPLEXITY ASSESSMENT OF G-CODE MODELS FOR PRINT QUEUE MANAGEMENT IN A 3D FARM

Queue management in 3D-printing farms depends on prior knowledge of the geometric complexity of the parts to be produced, since complex parts, with pronounced overhangs, disconnected islands, or organic surfaces, increase filament retractions, printing time, and failure risk. Existing complexity indices, however, are computed from the CAD or STL model, unavailable at the scheduling decision point, where only the sliced G-code exists. This paper addresses that gap with a complexity criterion computable exclusively from the G-code, extending Forrest's taxonomy with a fourth, manufacturing dimension capturing non-trivial slicer events such as retractions and bridges. Seven numerical attributes are extracted by a purpose-built parser and combined into a composite score, which defines a complex/non-complex screening rule via percentile thresholds. A Random Forest classifier, chosen for its robustness to correlated attributes and low computational cost, is trained to reproduce this rule in real time directly from the raw attributes. Evaluated on 1,745 G-codes from 50 ModelNet categories, the classifier reaches an F1-score of 0.991, ROC-AUC of 0.999, in 5-fold cross-validation, 0.989 on a 175-sample test set, and correctly classifies all 6 anchor parts with confidence above 97\%. These results confirm the criterion is fully operational from G-code alone, offering an objective basis for complexity screening in 3D Farm print queues; because the criterion and its labels share the same attributes, the metrics measure fidelity to this engineered rule, and validating it against independent proxies such as measured print time is discussed as a direction for future work. The approach contributes a practical, deployable estimator for parallel-machine scheduling in Industry 4.0 environments.