The growing demand for precision channel fabrication in difficult-to-machine materials such as Inconel necessitates reliable prediction of AWJ milling characteristics. Conventional data-driven models often lack generalization under unseen conditions, leading to inconsistent performance. To address this limitation, physics-guided and conventional multi-layer perceptron (MLP) regression models were developed, and their performance was evaluated under unseen input data conditions for predicting surface roughness (Ra), milling depth, and kerf angle. Experiments were conducted on Inconel 718 using a Box–Behnken design within the Response Surface Methodology framework, considering abrasive mesh size (80 and 120), abrasive flow rate (AFR), water pressure, and traverse rate. The underlying physical mechanisms were analysed, and ANOVA was used to identify significant features. Physics-based insights were incorporated for feature prioritization in the physics-guided MLP regression models, while the conventional MLP regression models were developed without such constraints. The models were trained using a bootstrap-resampled dataset with normalized features, and hyperparameters were optimized using GridSearchCV with five-fold cross-validation to ensure a fair comparison. For 80 and 120 mesh abrasives, minimum Ra values of 0.477 µm and 0.472 µm were obtained at a water pressure of 175 MPa, traverse rates of 1600 mm/min and 1300 mm/min, and AFR values of 0.42 kg/min and 0.32 kg/min, respectively. Milling depth was primarily influenced by water pressure and traverse rate, while kerf angle was governed by abrasive size and AFR. Under unseen conditions, the physics-guided model achieved lower prediction errors of 3.3–4.1% for milling depth, 1.75–2.4% for kerf angle, and 3.6–4.3% for Ra, compared to significantly higher errors of 30.3–32.8%, 13.0–14.8%, and 21.4–49.4% for the non-physics-guided model. Overall, the integration of physics-based feature prioritization with optimized MLP regression models provides a robust and industrially applicable framework for predicting AWJ channel milling characteristics, thereby reducing wastage of costly workpiece materials, minimizing machine resource usage, and lowering production costs.
[Display omitted]
- Physics-guided artificial neural network model for predicting abrasive waterjet channel milling characteristics in Inconel
- Prabhu Ramesh - Anna University, ChennaiCyril Garg - Delhi Private School, Sharjah, United Arab EmiratesYakub Mogul - University of Greater ManchesterJaimon Dennis Quadros - University of Greater Manchester, National Centre for Motorsport EngineeringGanesan Subramanian
- Materials today communications, Vol.54, 115640
- Elsevier Ltd
- Data will be made available on request.
- 9979934308841; 2352-4928; 2352-4928; 2352-4928
- © 2026 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/bync-nd/4.0/).
- National Centre for Motorsport Engineering
- English
- Journal article
- 29/06/2026
- 28/06/2026