Accelerating process development for 3D printing of new metal alloys

Additive manufacturing (AM), particularly laser powder bed fusion (L-PBF), is revolutionizing various industries. However, variability in the quality of 3D-printed metal parts, affecting mechanical properties and dimensional accuracy, hinders widespread adoption. This variability stems from factors including powder and machine variations, scanning strategies, and printing conditions. Current process development relies heavily on ex situ methods, which are time-consuming and inefficient. In situ approaches, while offering real-time monitoring, are often limited by observable features, algorithmic accuracy, and the need for complex, high-cost setups. This research aims to overcome these limitations by developing an efficient and generalizable in situ method for process development in 3D metal printing, focusing on achieving melt-pool stability and addressing the variability problem.

Prior research has explored the impact of process parameters on the quality of 3D-printed metal parts. Studies have investigated the relationship between laser beam power and velocity, energy density, molten pool stability, and the occurrence of defects such as lack-of-fusion, keyholing, and balling. High-speed imaging has been employed to monitor molten pools, but its application in in situ process mapping has been hampered by limited observable features, low algorithmic accuracy, and the requirement for complex and expensive temperature measurement setups. Existing methods often rely on convolutional neural networks (CNNs) for image analysis, but the need for high accuracy and generalizability motivates the exploration of alternative approaches.

This study proposes an in situ approach using a novel combination of high-speed imaging and video vision transformers. An off-axial imaging setup, designed to minimize interference from the plasma plume, captures melt-pool dynamics at 54,000 frames per second. The high temporal resolution allows for detailed analysis of melt-pool oscillations. The captured videos are processed using image registration and intensity normalization. The core of the methodology involves using video vision transformers (ViViTs) to classify the printing process into four categories: desirable regimes and three types of defects (keyholing, balling, and lack-of-fusion). ViViTs, leveraging temporal embedding, provide superior accuracy compared to CNNs and traditional computer vision approaches. Regularization techniques are employed to handle the limited data available from the high-speed camera. The model's generalizability is tested by performing cross-dataset evaluations across different metal alloys (stainless steel SS316L, titanium alloy Ti-6Al-4V, and Inconel alloy IN718) with varying thermophysical properties. The method generates process maps, showing the relationships between laser power, scan velocity, and printing outcomes. In addition, process maps for morphological variability (melt pool width and area) are generated to further enhance process optimization.

The developed in situ method achieves high classification accuracy (over 98%) in identifying printing regimes and defects. Cross-dataset evaluations demonstrate the method's generalizability across different alloys, with top-1 accuracy reaching 98%. Process maps generated using the ViViT model accurately predict printability and defect formation across the tested alloys, largely matching ex situ characterization results. Analysis of melt-pool dynamics reveals that keyholing regimes exhibit high-frequency intensity fluctuations (8-17 kHz), while stable regimes show more stable intensity values. Variability maps indicate that higher energy densities lead to larger standard deviations in melt-pool width and area, highlighting the importance of process parameter selection for consistent printing quality. Comparisons with other state-of-the-art models (VGG16, ResNet152, MoViNet-A1, TimeSformer) confirm the superior performance of video vision transformers. Ablation studies demonstrate the importance of sufficient temporal context and regularization for optimal performance. The t-SNE projection of ViViT-B features and attention maps showcase clear separation between melt-pool classes and the importance of dynamic features in the keyhole and tail regions.

The findings demonstrate the feasibility and effectiveness of using a simple off-axial imaging setup in conjunction with video vision transformers for in situ process monitoring and process map generation. The high accuracy and generalizability of the method provide a powerful tool for accelerating process development for new metal alloys. The incorporation of temporal features is crucial for accurately predicting printing regimes, as the dynamics of the process significantly contribute to defect formation. The ability to identify defects in situ, without the need for complex and costly setups, represents a significant advance in additive manufacturing process development. The generated process and variability maps guide the selection of optimal process parameters for melt-pool stability and consistent part quality. The high performance of the video vision transformer model highlights its potential as a robust and reliable tool for real-time process control in the future.

This study presents a novel in situ method for accelerating process development in 3D metal printing. The method utilizes high-speed imaging and video vision transformers to achieve high-accuracy process mapping and defect detection across different alloys. The results demonstrate the potential to significantly reduce the time and cost associated with developing new metal alloy printing processes. Future work will focus on addressing the challenges of real-time monitoring and integrating this method into closed-loop control systems for fully automated process optimization.

While the method demonstrates strong generalizability, limitations exist. The accuracy is somewhat affected by the vapor plume density, particularly with Ti-6Al-4V. Real-time implementation requires higher data transfer rates and longer video recording capabilities from high-speed cameras. Further research is needed to fully address these limitations and to develop multi-label classification models to better distinguish between subtly different defect types.