An augmented surprise-guided sequential learning framework for predicting the melt pool geometry

Ahmed Shoyeb Raihan, Hamed Khosravi, Tanveer Hossain Bhuiyan, Imtiaz Ahmed

Abstract: Metal Additive Manufacturing (MAM) has transformed the manufacturing landscape, bringing notable benefits such as intricate design capabilities, minimal material wastage, rapid prototyping, compatibility with diverse materials, and customized solutions. However, the complete adoption of this technology in the industry is impeded by challenges in ensuring uniform product quality. A pivotal aspect of MAM’s successful implementation lies in understanding the intricate relationship between its process parameters and the melt pool characteristics. In this scenario, the integration of Artificial Intelligence (AI) into MAM is vital. While traditional machine learning (ML) approaches are effective, they typically rely on large datasets to accurately capture complex relationships. However, in MAM, creating such extensive datasets is both time-consuming and resource-intensive, posing a significant challenge for effectively applying these methods. Our study addresses this challenge by introducing a novel surprise-guided sequential learning framework (SurpriseAF-BO). The framework represents a paradigm shift in the field of MAM, leveraging an iterative and adaptive learning approach. It efficiently models the dynamics between process parameters and melt pool characteristics with limited data which is a critical advantage in the cyber manufacturing environment of MAM. In comparison to conventional ML models, our sequential learning method demonstrates superior predictive accuracy for melt pool depth, width, and length. To further improve our methodology, we have incorporated a Conditional Tabular Generative Adversarial Network (CTGAN) model into our framework. It produces synthetic data that closely mirrors real experimental data, enhancing the overall learning process. We call this improved framework ‘CT-SurpriseAF-BO’. This advancement in our sequential learning approach significantly strengthens its predictive accuracy, all while eliminating the need for extra physical experiments. Additionally, our frameworks’ versatility and effectiveness are validated through their application to two other distinct datasets in material science and manufacturing, demonstrating their broad applicability. Our study not only showcases the application of cutting-edge data-driven techniques but also highlights the significant impact of sequential AI and ML in the field of cyber manufacturing, especially in MAM.