A new learning framework combines incremental learning with transfer learning to predict energy consumption in directed energy deposition during metal additive manufacturing, boosting accuracy and generalizability with limited data.

Published in npj Advanced Manufacturing, the approach enables robust, real-time predictions and cross-scenario generalization across different materials and process conditions.

The neural network design uses backpropagation, dropout, and batch normalization; memory buffers and meta-learning support continual knowledge retention, while domain-adaptive transfer learning aligns feature representations to minimize distributional differences.

The study draws on data from 20 samples of CoCrMo or IN718 to conduct three transfer tasks, transferring knowledge across materials and processing conditions.

While promising, the work notes generalizability limits under substantial material or process changes and points to future directions like GAN-based data augmentation and physics-informed simulations to enrich training data.

Future work includes expanding to more materials and processes and integrating energy forecasts with mechanical properties and surface quality to create a holistic optimization platform.

Diverse DED datasets with variables such as laser power, scan speed, and powder feed rate were used, achieving notable improvements in prediction accuracy and resilience to catastrophic forgetting via memory replay buffers and meta-learning.

The research highlights environmental and workforce benefits, emphasizing energy-efficient manufacturing and the augmentation of human expertise with predictive analytics.

Incremental learning enables continuous model updates as new data arrive without full retraining, while transfer learning facilitates knowledge transfer across related DED scenarios to reduce labeled data needs and speed learning.

Across tasks, models that use incremental learning integrated with transfer learning outperform vanilla models, with LSTM and TCN offering the strongest predictive performance and transformers showing notable gains in certain scenarios.

The methodology follows a three-stage workflow: source-domain incremental pre-training, target-domain re-training, and target-domain testing, with explicit updates to model parameters via incremental learning.

Material transfer across alloys induces greater domain shift than transferring laser power within the same material, yet incremental learning narrows the resulting performance gap.