Milling Surface Roughness Prediction Based on Physics-Informed Machine Learning.

Area of Science:

• Mechanical Engineering
• Materials Science
• Artificial Intelligence

Background:

• Surface roughness is critical for mechanical product performance, influencing fatigue strength and wear resistance.
• Existing machine learning models for surface roughness prediction often converge to local minima, leading to poor generalization and physical law violations.

Purpose of the Study:

• To develop a physics-informed deep learning (PIDL) method for milling surface roughness prediction.
• To enhance model generalization and ensure predictions adhere to physical laws.

Main Methods:

• Integrated physical knowledge into the input and training phases of deep learning.
• Employed data augmentation using surface roughness mechanism models.
• Utilized a CNN-GRU architecture with bi-directional GRUs and multi-headed self-attention.
• Developed a physically guided loss function for model training.

Main Results:

• The PIDL model achieved the highest prediction accuracy on the S45C and GAMHE 5.0 datasets.
• Reduced mean absolute percentage error by an average of 3.029% compared to state-of-the-art methods.
• Demonstrated superior performance in milling surface roughness predictions.

Conclusions:

• Physics-informed deep learning offers a promising approach for accurate and reliable surface roughness prediction.
• Integrating physical constraints into machine learning models enhances their generalization and predictive power.
• This methodology represents a potential future direction for machine learning evolution in engineering applications.