SK hynix | Digital Transformation (DT), Manufacturing Execution System (MES)
Iljeok Kim is currently a research engineer, developing automated material handling system (AMHS) at SK hynix. I received a Ph.D. degree in mechanical engineering from Industrial AI Lab at Pohang University of Science and Technology (POSTECH), under the supervision of Prof. Seungchul Lee (Currently Korea Advanced Institute of Science and Technology, KAIST). He also served as a research scientist at the AI Research Center for Manufacturing Systems at Korea Institute of Industrial Technology (KITECH), under the guidance of Ph.D. Jong Pil Yun. His primary research interests span AI-enabled smart manufacturing, physics-integrated AI, and trustworthy AI in industrial applications.
The overarching objective of my research is to pioneer how to integrate and interact physics knowledge and AI to understand the dynamic behavior of a system or process. In accordance with the combining method of principle knowledge, we aim to develop mathematical models for dynamic systems: implicit knowledge integration in AI, explicit knowledge integration in AI, and knowledge discovery from AI.
Additionally, I am also interested in research to ensure robust and reliable deep networks, including debiased networks, uncertainty quantification, and domain adaptation (generalization).
Github
Implicit knowledge integration in AI has been proposed to implicitly embed physics knowledge in deep networks. This study illustrates Radon transform-integrated NN to alleviate the inverse problem in CT reconstruction. 1) Coordinate sampling along the X-ray path to represent the system geometry. 2) Radon transform-integrated loss design for inferring attenuation coefficients based on measured sinograms. 3) Space-variant Fourier encoding for high-frequency product defect representation to overcome spectral bias of deep networks. 4) Bin scheduler for accelerating deep network training based on spectral bias. 5) 3D rendering visualization based on various object functions without additional engines by combining rendering techniques based on learned neural attenuation fields.
Explicit knowledge integration in AI has been proposed to explicitly combine principle knowledge in deep networks. Data distribution shifts must be addressed to construct robust and reliable deep networks. However, it is difficult to solve the domain shift problem using a learning methodology due to the characteristics of maximum likelihood estimation (MLE). Therefore, this study aims to represent data acquired under various operating conditions into domain invariant mapping by signal-preprocessing for machine health management. 1) Signal-preprocessing for domain generalization in time and frequency domains. 2) Signal-preprocessing for domain interpretation for extracting and visualizing fault signals. 3) Physically interpretable XAI representation for reliable inference validation on untrained domains. Furthermore, additional research on simulation modeling for each fault type is being conducted to learn deep networks in the absence of any fault signals.
Knowledge discovery from AI has been proposed to extract information from data that is difficult to analyze with domain knowledge. The injection molding process is a complex process involving multi-physics phenomena, making it very difficult to implement a process optimization model because it is difficult to determine the importance of features. Thus, we propose a novel method to directly infer feature importance using the trained weights of the deep network. 1) SHAP and LIME, which have been most widely used in trustworthy AI, have the limitation that the inference model and the interpretation model are separated. 2) The existing LRP decomposes relevance scores by utilizing the internal structure of deep networks for classification evidence. 3) The proposed LRP-R approximates the learned weights with the regression coefficients of the features by defining three constraints. 4) Inference and interpretation are performed simultaneously based on the learned weights, allowing for faster and more accurate feature importance scores to be estimated.
Additionally, existing XAI is very vulnerable to spurious correlation due to MLE learning mechanism. For example, when trying to predict drum class, humans who appear intermittently also have high relevance scores. The proposed method combines XAI and bayesian deep networks, aiming to infer uncertainty maps. This method can provide additional information about spurious correlations for classes from an epistemic uncertainty perspective. When comparing before and after inpainting of medical marks using a conditional diffusion model, before inpainting, not only is the focus on marks high, but uncertainty is also visualized. On the other hand, after inpainting, it can be confirmed that the features corresponding to the lesion are extracted accurately without uncertainty. In other words, this method aims to extract additional information from the uncertainty perspective beyond trustworthy AI from the MLE and posterior mean perspectives. This research was also applied to industrial data consisting of structured data.
Industrial AI Lab at Pohang University of Science and Technology (POSTECH)
AI Research Center for Manufacturing Systems (AIMS) at Korea Institute of Industrial Technology (KITECH)
AI Research Center for Manufacturing Systems (AIMS)
Ph.D. Proposals at Pohang University of Science and Technology (POSTECH)
Email: [email protected] / [email protected]