Session: 03-04-02: Artificial Intelligent Applications in Manufacturing II

Paper Number: 173918

Robust Multimodal Learning of Melt Pool Dynamics in Laser Powder Bed Fusion 

Advanced manufacturing machines, such as laser powder bed fusion (LPBF) systems, rely on various sensors for in-situ process monitoring to ensure build quality and consistency. However, the signals obtained from these sensors are often noisy, making it difficult to extract meaningful insights from a single data modality. Accurately capturing and correlating melt pool dynamics requires a more robust approach that integrates multiple data sources.

In this study, we utilized the 2022 NIST AM Bench asynchronous challenge dataset, which provides two complementary data modalities: X-ray images and laser absorptivity data. Each modality offers unique insights into the melt pool behavior, making them valuable for understanding the complex thermal and fluid dynamics of the LPBF process. To analyze these modalities, we employed a U-Net-based Convolutional Neural Network (CNN) and a Long Short-Term Memory (LSTM) Recurrent Neural Network (RNN) to process each modality separately. The CNN architecture effectively captures spatial features from X-ray images, while the LSTM model learns temporal dependencies from laser absorptivity data. We developed a mask-based image segmentation approach to automate the melt pool and keyhole detection from the NIST AM Bench dataset.

To further enhance the predictive capability and robustness of our approach, we concurrently trained both modalities using a Conv-LSTM model. This hybrid model integrates CNN and RNN architectures, allowing it to extract both spatial and temporal features while learning correlations between the two data sources. The multimodal fusion approach significantly improves training stability and model accuracy.  Further, a non-dimensional parameter of normalized energy density is used as a feature to train the models and make the prediction material-agnostic.

A key strength of the proposed framework lies in its ability to maintain effective performance even in the absence of one data modality. By adopting a modality-agnostic design, the model ensures robustness under varying sensor conditions, such as X-ray imaging or laser absorptivity data. This adaptability is particularly advantageous for real-world applications in Laser Powder Bed Fusion (LPBF), where sensor availability and quality may fluctuate. Consequently, this capability enhances the reliability of process monitoring and defect detection, thereby contributing to improved manufacturing consistency and quality control. 

Moreover, the framework can be extended to incorporate additional sensor modalities, including thermal imaging, acoustic emissions, and high-speed video data. The integration of these diverse data sources would enable a more comprehensive analysis of melt pool dynamics, allowing the model to function as an advanced tool for LPBF process fingerprinting. This, in turn, facilitates real-time monitoring and adaptive process control, further strengthening defect mitigation strategies.  By leveraging robust multimodal learning, this approach supports the advancement of intelligent manufacturing, promoting higher precision and reliability in additive manufacturing processes.

Presenting Author: Satyajit Mojumder Washington State University

Presenting Author Biography: Satyajit Mojumder is a tenure-track assistant professor in the School of Mechanical and Materials Engineering at Washington State University in Pullman. He holds a PhD in Theoretical and Applied Mechanics from Northwestern University. His research focuses on innovative computational methods for advanced materials and manufacturing systems, integrating data science algorithms to create mechanistic reduced-order models for complex computational challenges. These efforts have led to the establishment of a startup, HIDENN-AI, LLC, and secured funding for multiple SBIR/STTR projects, including NSF ACCESS computational resources, for which he was the Principal Investigator. He is an active member of ASME and USACM and serves as a referee for 20 journals in computational mechanics and related fields. Satyajit earned his Bachelor’s (2015) and Master’s (2017) degrees in mechanical engineering from Bangladesh University of Engineering and Technology (BUET) in Dhaka. He previously served as a faculty member at BUET for three years. His collaborative research has resulted in 43 journal articles, 17 conference proceedings, 2 invited talks, 4 patents, and various fellowship grants and awards, including multiple NSF travel grants.

Authors:

Satyajit Mojumder Washington State University
Tiana Tonge Washington State University
Pallock Halder Washington State University