Tackling the ‘small data’ challenge: A Versatile Melt Pool Encoder via Reconstructive Self-supervised Pretraining for Few-Shot Learning

Research in metal additive manufacturing frequently encounters ‘small data’ challenges due to the high costs associated with data generation and the extensive categorization of the process. In addition, most of the available data is unlabeled, i.e., it is not data science-ready. In this paper, we demonstrate a method that overcomes this issue for analyzing melt pool (MP) signatures, a key real-time indicator of process health. Specifically, we propose a novel approach that utilizes reconstructive self-supervised pretraining to develop a pre-trained MP encoder that is subsequently finetuned for multiple downstream tasks—–including predictions of MP width, aspect ratio, size, and average temperature—–using only a few labeled samples. By leveraging dense unlabeled visual monitoring data, our method significantly improves prediction accuracy compared to traditional end-to-end supervised learning, especially in few-shot scenarios. Comprehensive computational experiments demonstrate that our pretrained encoder, when finetuned with a simple regression head, achieves a superior performance while reducing the overall reliance on labeled data. It thus offers a cost-effective solution for advanced manufacturing applications, e.g., real-time process control.