Deep learning based heat transfer simulation of the casting process.

Area of Science:

• Materials Science
• Computational Science
• Artificial Intelligence

Background:

• Numerical simulations for casting solidification are computationally intensive and time-consuming.
• Existing methods often require complex constitutional models.
• Predicting temperature fields accurately is crucial for optimizing casting processes.

Purpose of the Study:

• To develop a rapid and accurate method for predicting temperature fields during casting solidification.
• To overcome the limitations of traditional numerical simulation techniques.
• To utilize deep learning for efficient thermal analysis in casting.

Main Methods:

• Developed modified U-net network architectures incorporating Inception and Convolutional Block Attention Module (CBAM) modules.
• Generated training data from 200 diverse geometric models with casting, mold, and chill components.
• Used Finite Difference Method (FDM) simulations to obtain temperature fields for training data.
• Trained deep learning models to predict temperature fields at time t_i+1 from input at time t_i.

Main Results:

• Achieved an average prediction accuracy of 94.5% with an absolute temperature error of 7 ℃.
• Demonstrated rapid prediction capabilities, with each time step prediction taking only one second.
• Successfully handled complex multi-component and multi-material geometries, including casting, chill, and mold.
• Models showed proficiency in forecasting temperature fields for arbitrarily shaped castings at various time points.

Conclusions:

• Deep learning models offer a swift and accurate alternative to traditional numerical simulations for casting solidification.
• The proposed U-net architecture with Inception and CBAM modules effectively predicts temperature fields in complex casting scenarios.
• This approach significantly reduces computational cost and time, enabling faster optimization of casting processes.