Session: CIE–31: Graduate Student Poster Symposium

Paper Number: 148505

148505 - Am Transformer: A Koopman Theory-Based Transformer to Learn Additive Manufacturing Dynamics 

Understanding and predicting melt pool dynamics are highly significant for Additive Manufacturing (AM) because of their immense impact on manufactured products' quality and mechanical properties. Recent studies using data-driven machine learning approaches show promising results. However, these approaches have limitations as they rely solely on data without considering the underlying dynamics. In this study, we propose a novel approach, AM Transformer, which helps to understand melt pool dynamics and predict its behaviors. The proposed method offers a fundamental deep-learning architecture to comprehend complex spatiotemporal dependencies among melt pools and predict future melt-pool behaviors. Our study introduces a new approach called AM Transformer, which can aid in comprehending the dynamics of melt pools and predicting their behaviors. This method employs a deep-learning architecture to understand the complex spatiotemporal dependencies among different melt pools. Based on this understanding, it can predict the future behaviors of melt pools. To improve the understanding of the melt-pool dynamics, our model adapts the Koopman operator to generate latent embeddings, which capture the relationships among the melt-pool state dynamical transitions. Thanks to the Transformer's attention mechanism, the AM Transformer can learn spatial and temporal dependencies among melt pools. Our case study shows that the AM Transformer outperforms the combination of a standard autoencoder with the Transformer model and ConvLSTM models regarding both criteria: the representation of the melt pool morphology and the congruence between predicted and actual melt pool dimensions. When evaluating the morphology, the AM Transformer shows an MAE of 3.6227 and SSIM of 0.9206. Regarding the melt pool size estimation, the AM Transformer has the lowest MAE of 0.0009 and an accuracy of 0.9273. This indicates the AM Transformer's improved robustness in predicting melt pool behavior. AM Transformer can also give us opportunities to explore the incorporation of a priori physical knowledge, enhancing our understanding of melt-pool dynamics.

Presenting Author: Suk Ki Lee Arizona State University

Presenting Author Biography: Suk Ki Lee is currently a PhD student at Arizona State University's School of Manufacturing Systems and Networks. Prior to this, he obtained his Bachelor's and Master's degrees in electronic and electrical engineering from Sungkyunkwan University, South Korea. After gaining over 8 years of experience as an engineer at Samsung Electronics, he decided to pursue research in advanced manufacturing. His research interests include machine learning and AI in manufacturing, data-driven optimization in system design, and additive manufacturing.

Authors:

Hyunwoong Ko Arizona State University
Suk Ki Lee Arizona State University