DeepMS: A data-driven approach to machining process sequencing using transformers

Abstract

Efficient and intelligent machining process sequencing remains a key challenge in computer-aided process planning (CAPP). Traditional methods often rely on manually defined rules and explicit feature recognition, limiting their adaptability across diverse parts and evolving manufacturing environments. Recent advances in deep learning (DL), particularly in transformer-based sequence modeling, offer a promising alternative by enabling systems to learn sequencing logic directly from data without explicitly modeling complex rules. This paper presents a novel DL framework that predicts machining sequences directly from the 3D geometry of final parts. Operating on voxelized representations, the model generates an ordered sequence of machining operations, each associated with a volumetric shape representing the material removed from raw stock—eliminating the need for predefined features or rule-based logic. The framework integrates a transformer-based sequence autoencoder to model operation order and an encoder based on 3D convolutional neural networks (CNN) to map final part geometry to sequence representations. To efficiently handle high-dimensional voxelized data, a 3D CNN autoencoder is employed to compress voxelized removal volumes. Components of these pretrained models are combined into an inference pipeline that generates machining sequences directly from the final part geometry. Trained on a synthetic dataset of 1.08 million prismatic parts with embedded geometric precedence rules, the framework achieves a sequence prediction accuracy of 99.48 % and reconstructs final part geometry with a volumetric intersection-over-union (IoU) of 97.33 %. Results show the framework can generalize sequencing logic and material removal volumes from geometry data alone, offering a flexible and scalable approach to process planning and laying the foundation for future extensions in real-world manufacturing scenarios.