Geometric Deep Learning for Milling Process Time Estimation: A Graph Neural Network Approach

In the manufacturing sector, which is becoming increasingly automated at all levels, the ability to obtain fast and accurate estimates of processing times is a key requirement for ensuring production efficiency. In milling, machining time estimation is typically derived from toolpath simulations, which, although precise, are time-consuming and strictly dependent on the specific machine setup and cutting strategy. This dependency limits their use in the early stages of production planning and limits flexible manufacturing. This work proposes a deep learning approach based on Graph Neural Networks (GNN) that directly exploits the geometric information contained in STL files to predict milling process times. Unlike traditional machine learning models that rely solely on tabular inputs, the proposed method combines geometric representations, encoded as graphs, with pre-computed numerical features. Through graph convolutional layers, the model learns a compact embedding of the part geometry, while a learnable process embedding captures process-specific information. These representations are then fused within fully connected layers to produce an accurate prediction of the machining time for each milling process. The model was evaluated on a real industrial dataset including multiple workpieces and tested using a Leave-One-Piece-Out (LOPO) cross-validation strategy. Its performance was compared with three classical machine learning methods, Random Forest, Decision Tree, and Support Vector Machine. The GNN showed superior accuracy and generalization, effectively estimating machining times for unseen workpieces. The results indicate that, with a general R² of 0.94, the proposed neural network can provide highly accurate machining time estimations that effectively support production planning from the earliest stages of the manufacturing process.