Physics-Informed Online Learning for Temperature Prediction in Metal AM.

Area of Science:

• Materials Science
• Mechanical Engineering
• Computational Science

Background:

• Precise temperature field prediction is vital for metal additive manufacturing (AM) process control and optimization.
• Traditional offline and data-driven methods lack real-time adaptability for dynamic AM scenarios.

Purpose of the Study:

• To introduce the first physics-informed (PI) online learning framework for real-time temperature prediction in metal AM.
• To address the limitations of traditional methods in adapting to new process conditions.

Main Methods:

• Development of a physics-informed neural network (PINN) integrating neural networks with physics-informed inputs and loss functions.
• A two-phase learning approach: initial pretraining on known data, followed by dynamic online weight updates with new data.
• Leveraging physical laws and prior manufacturing knowledge within the PINN framework.

Main Results:

• The PI online learning framework accurately predicts temperature fields for unseen metal AM processes.
• Demonstrated superior performance over traditional data-driven models, particularly in predicting the Heat Affected Zone (HAZ) and melt pool.
• Identified the impact of hyperparameters like learning rate and batch size on online learning optimization.

Conclusions:

• The proposed PINN-based framework offers robust and accurate real-time temperature prediction for metal AM.
• This approach significantly enhances adaptability to diverse process parameters, geometries, and materials.
• The framework holds substantial potential for improving online control and optimization in metal AM.