Improving five-axis milling force prediction through physics-guided transfer learning and orientation-aware process planning

Five-axis milling is widely used for manufacturing complex-geometry components in aerospace, energy, die and mold, and biomedical applications, where accurate cutting-force prediction is essential for process planning, optimization, and digital-twin development. However, mechanistic models require material- and condition-specific calibration, while purely data-driven approaches often demand extensive experimental data and may generalize poorly beyond the training domain. This study proposes the Physics-Guided Milling Network (PGMNet), a simulation-to-experiment transfer-learning framework that combines mechanistic modeling with deep learning. The network is pretrained on a large mechanistic simulation dataset and fine-tuned using limited experimental force measurements. Process parameters, tool geometry, tool orientation, and thermo-mechanical material properties are used as inputs to enable transferable force prediction across materials and cutting conditions. For AISI 1050, Al 7075-T6, and Ti-6Al-4V, PGMNet achieved R 2 values above 98% with RMSE below 20 N, outperforming the mechanistic model and conventional regression baselines. For unseen Inconel 625, mean R 2 remained above 91%, while RMSE was reduced by up to 45% relative to the mechanistic simulation. Lead–tilt force maps further demonstrated the framework's potential for orientation-aware five-axis process planning.