Machine Learning Enhances Additive Manufacturing: Predicting Laser Absorption in Real Time

By Soham NandiReviewed by Susha Cheriyedath, M.Sc.Jan 10 2024

In an article published in the journal Nature, researchers explored quantifying absorbed light in laser-material interactions for additive manufacturing (AM). Employing in situ high-speed synchrotron X-ray visualization and integrating sphere radiometry, they proposed two approaches. The end-to-end method utilized deep convolutional neural networks (ConvNets) to predict laser absorptance from X-ray images. The two-stage approach employed semantic segmentation for geometric features, coupled with classical regression models.

Background

Keyholing, observed in welding and metal AM, involves vapor depression formation due to high laser power. The vapor depression shape affects laser absorptance, influencing fabrication efficiency. Existing models, often complex or limited by software access, pose challenges. Analytical models, like the Rosenthal equation, provide estimates for conduction-mode melting but falter in keyhole mode. High-fidelity multiphysics models offer improved accuracy for deep keyholes but are limited by software accessibility.

Experimental methods like calorimetry and integrating spheres measure average or real-time absorptance but face limitations in accessibility and setup costs. To overcome these limitations, the research established a temporally resolved vapor depression-absorption dataset using synchrotron X-ray imaging and applied machine learning (ML) models for absorptance prediction. This ML approach, reliant on geometric vapor depression features, reduced the need for costly experiments and enhanced the connection between industrial parameters and predictive models.

The study introduced two prediction approaches: an end-to-end method employing ConvNets and a modular approach combining semantic segmentation and regression models. Both methods demonstrated low mean absolute error (MAE), offering efficient alternatives for real-time monitoring in metal AM processes.

Methods

The researchers utilized data from the synchrotron X-ray imaging system at the advanced photon source (APS) for ML model development in laser energy absorptance prediction and vapor depression segmentation. A ytterbium fiber laser with a wavelength of 1070 nm and a maximum power of 540 W, along with a galvanometer scanner, enabled laser scanning across samples in a stainless steel chamber with a 1 atm Ar environment.

High-speed X-ray imaging at 50 kHz captured vapor depression images during laser-sample interactions. Laser energy absorptance was measured using a calibrated integrating sphere, with the setup designed to capture strong backscattered light. The dataset included images with and without a powder layer.

For ML, two ConvNets models, ResNets and ConvNeXt, were trained on the absorptance dataset. Training configurations included both randomly initialized weights and pre-trained weights from ImageNet. Segmentation models, employing UNet, UNet++, DeepLabV3, and DeepLabV3+, were trained for vapor depression image segmentation. The training strategy involved data augmentation for diverse learning. The RMSprop optimizer, binary cross-entropy, and dice loss were used, with a learning rate decay strategy and early stop mechanism.

Geometric features were extracted for the depression beneath the substrate surface, encompassing area, depth, and various width measurements. The authors aimed to predict laser energy absorptance and perform segmentation based on geometric features. The research contributed by establishing benchmark datasets, exploring ML approaches, and revealing correlations between geometric characteristics and laser energy absorption, offering an experiment-validated laser absorption model for Ti-6Al-4V.

Results

The study presented extensive datasets and models for predicting laser energy absorptance and segmenting vapor depressions in metal AM. Two datasets were compiled, one without a powder layer and another with a 100 μm Ti64 powder layer. The end-to-end approach, employing ResNet-50 and ConvNeXt-T models, demonstrated effective predictions of laser absorptance, with ConvNeXt-T exhibiting superior performance. Notably, ImageNet pre-trained weights significantly impacted ConvNeXt-T, while ResNet-50 showed a milder effect. The models successfully captured different stages of laser-metal interaction, including keyhole formation.

For the modular approach, UNet achieved optimal performance among segmentation models, effectively delineating vapor depressions. The subsequent use of geometric features extracted from the segmented masks facilitated laser absorptance prediction. Linear regression, decision tree, random forest, and extreme gradient boosting (XGBoost) models were compared, with random forest yielding the best results on the test dataset. The modular approach's performance was contingent on successful image segmentation and feature extraction, presenting challenges in scenarios without a formed vapor depression.

Discussion

This discussion encompasses four key aspects: the impact of ImageNet pre-trained weights, model interpretability, the efficacy of fine-tuning with a powder layer, and practical recommendations. The researchers highlighted the significant influence of pre-trained weights from ImageNet on ConvNets' convergence speed and performance, especially notable in the ConvNeXt-T model. The model interpretation was addressed through techniques like Grad-CAM, highlighting ConvNeXt-T's gradual focus on relevant regions during different stages. Interpretability was crucial for understanding model decisions, especially in critical applications.

Fine-tuning with a powder layer was explored, demonstrating improved generalizability by leveraging prior knowledge from the dataset without powder. The modular approach, emphasizing interpretability through geometric features and random forest models, was contrasted with the end-to-end approach. While the latter excelled in operando applications prioritizing accuracy and automation, the former stood out when interpretability influenced decision-making.

Challenges in generalization, especially with powder involvement, were acknowledged, emphasizing the need for fine-tuning and domain-specific training data. Practical recommendations suggested prioritizing the end-to-end approach for operational settings and reserving the modular approach for applications where interpretability is paramount.

Conclusion

In conclusion, the research successfully addressed challenges in laser-material interactions for AM, offering efficient ML-based approaches for predicting laser absorptance and segmenting vapor depressions. The end-to-end and modular methods demonstrated effectiveness, balancing accuracy and interpretability. The contributions include benchmark datasets, insights into pre-trained weights' impact, and practical recommendations for diverse AM scenarios, highlighting the potential of ML in advancing real-time monitoring and decision-making in industrial processes.