A Physics-Informed Neural Network framework with strong robustness to low-accuracy physical models for predicting adhesive wear of self-made BNNC milling tool

TY - JOUR

T1 - A Physics-Informed Neural Network framework with strong robustness to low-accuracy physical models for predicting adhesive wear of self-made BNNC milling tool

AU - Tan, Shengyue

AU - Wang, Dongqian

AU - Liu, Yongliang

AU - Cai, Yonglin

AU - Wei, Jia

AU - Wang, Lei

AU - Teicher, Uwe

AU - Hänel, Albrecht

AU - Ihlenfeldt, Steffen

AU - Liang, Zhiqiang

N1 - Publisher Copyright: © 2026 The Society of Manufacturing Engineers

PY - 2026/2/28

Y1 - 2026/2/28

N2 - In high-speed hard milling, flank wear prediction of ultra-hard tools is necessary, where accuracy and stability are the two key indicators. Physics-Informed Neural Network (PINN) improves the prediction stability by embedding consistent physical laws into the training process. However, when low-accuracy physical models are commonly employed to constrain the solution space, optimization paths may be misled. This limits the performance of PINNs and their weighted frameworks. To address the issue of low-accuracy physical model misleading the optimization direction, a novel framework termed Physics-Informed Weighted Neural Network based on Prediction Error of Physics-driven models (PIWNN-PEP) is proposed. PIWNN-PEP can enhance the robustness of PINN against low-accuracy models. Furthermore, to capture long-term and complex dependencies over long time scales, a collaborative network xLSTM-Informer (xICNet) with a stacked mLSTM-sLSTM-Informer architecture is established, of which xICNet directly builds a mapping between multidimensional cutting forces and wear values within the PIWNN-PEP framework. The experimental result demonstrates that the proposed method prominently enhances robustness to low-accuracy physical models, compared with existing weighted PINN frameworks. The average tool wear prediction error remains below 1 μm.

AB - In high-speed hard milling, flank wear prediction of ultra-hard tools is necessary, where accuracy and stability are the two key indicators. Physics-Informed Neural Network (PINN) improves the prediction stability by embedding consistent physical laws into the training process. However, when low-accuracy physical models are commonly employed to constrain the solution space, optimization paths may be misled. This limits the performance of PINNs and their weighted frameworks. To address the issue of low-accuracy physical model misleading the optimization direction, a novel framework termed Physics-Informed Weighted Neural Network based on Prediction Error of Physics-driven models (PIWNN-PEP) is proposed. PIWNN-PEP can enhance the robustness of PINN against low-accuracy models. Furthermore, to capture long-term and complex dependencies over long time scales, a collaborative network xLSTM-Informer (xICNet) with a stacked mLSTM-sLSTM-Informer architecture is established, of which xICNet directly builds a mapping between multidimensional cutting forces and wear values within the PIWNN-PEP framework. The experimental result demonstrates that the proposed method prominently enhances robustness to low-accuracy physical models, compared with existing weighted PINN frameworks. The average tool wear prediction error remains below 1 μm.

KW - Binderless nanopolycrystalline cubic boron nitride (BNNC) ulter-hard tool

KW - High-speed hard milling

KW - Physics-informed neural network (PINN)

KW - Ulter-hard tool wear prediction

KW - xLSTM-informer

UR - https://www.scopus.com/pages/publications/105027882589

U2 - 10.1016/j.jmapro.2026.01.049

DO - 10.1016/j.jmapro.2026.01.049

M3 - Article

AN - SCOPUS:105027882589

SN - 1526-6125

VL - 160

SP - 285

EP - 301

JO - Journal of Manufacturing Processes

JF - Journal of Manufacturing Processes

ER -