AutoSplit: a two-stage AI architecture for enhanced classification of manufacturing processes with a focus on the identification of additive manufacturing components

In the product development phase of mechanical assemblies, engineers encounter an increasing variety of potential manufacturing routes for metal parts. Despite the advantages of additive manufacturing (AM), conventional methods often dominate due to a lack of interdisciplinary knowledge required for additive or hybrid manufacturing approaches. To streamline the development of hybrid manufactured components, this paper presents a novel two-stage methodology for automating part classification in manufacturing processes. A two-stage classification approach was selected to filter standard parts (e.g., screws, nuts, bolts), enabling a pre-filtering step that improves classification performance and reduces overfitting by minimizing the number of ST-components with similar features. The first stage employs convolutional neural networks (CNNs) for image-based classification and multi-layer perceptrons (MLPs) for feature-based classification, achieving 88.84% ±0.6 (SD) accuracy in differentiating standard from non-standard parts. The second stage utilizes a random forest classifier to categorize non-standard parts into three manufacturing processes (AM, machining, and sheet metal), achieving 82.0% ±1.1 (SD) accuracy, with particularly strong performance in machining identification (F1-score: 0.85 ±0.03 (SD)). The approach is trained on a comprehensive dataset of 20,000 CAD files sourced from GrabCAD, Fusion360, and TraceParts, evenly distributed across four categories. System performance was evaluated using fivefold cross-validation, demonstrating robust generalization across diverse part geometries and materials. This methodology provides guidance for selecting appropriate manufacturing routes for both redesigns and new designs.