Phy-Taylor: Partially Physics-Knowledge-Enhanced Deep Neural Networks via NN Editing

Area of Science:

• Artificial Intelligence
• Physics-Informed Machine Learning
• Engineering Applications

Background:

• Data-driven deep neural networks (DNNs) risk violating physical laws in engineering, causing unpredictable outcomes.
• Integrating domain knowledge is crucial for reliable AI in physical systems.

Purpose of the Study:

• To introduce Phy-Taylor, a physics-knowledge-enhanced DNN framework.
• To develop a method for accelerating the learning of physics-compliant representations.

Main Methods:

• Introduced a physics-compatible neural network (PhN) architecture using Taylor series monomials and noise suppressors.
• Developed a physics-guided neural network (NN) editing mechanism to enforce physics knowledge.
• Proposed a self-correcting Phy-Taylor extension for safety-critical autonomous systems.

Main Results:

• Phy-Taylor significantly reduces model parameters compared to traditional DNNs.
• Achieved accelerated training processes while maintaining high accuracy.
• Demonstrated enhanced model robustness and reliability in physical engineering tasks.

Conclusions:

• Phy-Taylor effectively integrates physics knowledge into DNNs, ensuring compliance with physical laws.
• The framework offers improved efficiency, accuracy, and robustness for AI in engineering.
• The self-correcting extension enhances safety for critical autonomous systems.