A welding penetration prediction model for laser welding process based on self-supervised learning using physics-informed neural networks

Researchers introduce SimPhysNet, a self-supervised learning algorithm that predicts laser welding penetration with 96.06% accuracy using only 200 labeled images—roughly 5% of typical datasets. The physics-informed neural network approach combines contrastive learning with few-shot learning to overcome the industrial manufacturing challenge of requiring extensive labeled data for quality assurance.

SimPhysNet addresses a persistent bottleneck in industrial automation: the scarcity of labeled training data for machine learning models in specialized manufacturing contexts. Laser welding penetration prediction is critical for defect-free joints, yet conventional supervised learning approaches demand thousands of meticulously labeled images, creating cost and resource barriers for manufacturers. The researchers solve this by embedding physical priors directly into a contrastive learning framework, allowing the model to extract meaningful features from unlabeled data while learning from minimal labeled samples through prototypical networks.

The technical innovation lies in combining three complementary approaches. Physics-informed neural networks constrain feature extraction to physically interpretable patterns in molten pools and keyholes, image augmentation improves generalization, and few-shot learning enables rapid deployment with limited labeled data. This contrasts sharply with standard deep learning workflows that struggle without large annotated datasets.

The impact extends beyond academic interest to manufacturing and industrial automation sectors seeking cost-effective quality control. For companies operating laser welding systems, the ability to achieve production-grade accuracy with 95% less labeling effort translates to reduced implementation costs and faster deployment timelines. The approach demonstrates a viable pathway for intelligent automation without prohibitive data collection expenses, potentially accelerating adoption of AI-driven quality assurance across precision manufacturing industries.

• →SimPhysNet achieves 96.06% classification accuracy using only 200 labeled images, comparable to supervised learning with full datasets.
• →Physics-informed neural networks embedded in contrastive learning frameworks enable meaningful feature extraction from unlabeled data.
• →Few-shot learning strategy based on prototypical networks enables robust classification with minimal labeled samples.
• →The approach reduces labeling requirements by 95%, lowering barriers to AI deployment in industrial manufacturing.
• →Method demonstrates practical pathway for intelligent automation in laser welding quality control without extensive data collection costs.