A Robust Recurrent Neural Networks-Based Surrogate Model for Thermal History and Melt Pool Characteristics in...

Area of Science:

• Additive Manufacturing
• Materials Science
• Computational Modeling

Background:

• Accurate control of melt pool characteristics is crucial in directed energy deposition (DED) for achieving desired material properties and geometric precision.
• Existing methods for predicting melt pool behavior can be computationally intensive and may lack real-time adaptability.

Purpose of the Study:

• To develop and validate a robust surrogate model for predicting melt pool characteristics in DED using recurrent neural network (RNN) architectures.
• To assess the predictive accuracy and computational efficiency of different RNN models, including Long Short-Term Memory (LSTM), Bidirectional LSTM (Bi-LSTM), and Gated Recurrent Unit (GRU).

Main Methods:

• Utilized a time series dataset generated from multi-physics simulations of the DED process.
• Employed a three-factor, three-level experimental design to capture a range of process conditions.
• Implemented and trained LSTM, Bi-LSTM, and GRU models to predict melt pool peak temperatures and geometric dimensions (length, width, depth).

Main Results:

• The Bi-LSTM model achieved a high R-square value of 0.983 for predicting melt pool peak temperatures.
• The GRU-based model demonstrated excellent performance for melt pool geometry prediction, with R-square values exceeding 0.88.
• RNN-based surrogate models reduced computation time by at least 29% compared to traditional methods, highlighting significant efficiency gains.

Conclusions:

• Recurrent neural network architectures provide a powerful and efficient tool for accurately predicting melt pool dynamics in DED.
• The developed RNN surrogate model enhances the understanding of melt pool behavior and facilitates precise setup and optimization of DED systems.
• This approach offers a pathway to improved quality control and process reliability in additive manufacturing.