Predicting energy consumption in directed energy deposition using incremental learning-integrated transfer learning

Metal additive manufacturing (AM) processes are often energy-intensive because of the use of high-energy heat sources. Predicting energy consumption accurately is critical for optimizing AM process parameters and minimizing environmental impact. Traditional machine learning models that predict energy consumption in metal AM processes are usually not generalizable when the material or process condition varies. To address this issue, we introduce an incremental learning-integrated transfer learning (TL) approach to predict energy consumption in the directed energy deposition (DED) process. Using a small dataset collected from 20 samples fabricated with CoCrMo or IN718, we conduct three TL tasks with varying process conditions. The incremental learning approach is integrated into the source domain pre-training step to learn knowledge from small datasets more efficiently. We evaluate the performance of the extreme gradient boosting (XGBoost), long short-term memory (LSTM), temporal convolutional networks (TCN), and transformer models. The TCN model achieves the best predictive performance with a mean absolute percentage error of 4.65%, a root mean squared error of 0.28, and a coefficient of determination of 0.92. The incremental learning-integrated TL framework achieves excellent predictive performance and generalizability with small volumes of data.