Author name cluster

Taehwan Kim

Possible papers associated with this exact author name in Arrow. This page groups case-insensitive exact name matches and is not a full identity disambiguation profile.

3 papers

1 author row

EAAI Journal 2026 Journal Article

Machine learning-aided network for process-property prediction of injection-molded polyamide-6 composite parts

Taehwan Kim
Unghyeon Cho
Jinwoo Jang
Min-Kyeom Kim
Yongjian Fang
Jin Young Jung
Jun Yeon Hwang
Sang Won Lee

Although process parameters significantly affect mechanical properties of injection-molded parts, optimizing them is challenging due to cost/time-intensive works for engineering applications. While design of experiments (DOE) is commonly used for process design, it still cannot fully address these challenges without understanding fundamental processing mechanisms. Therefore, this study proposes a machine learning-aided process-property prediction network (PPN), enabling reliable virtual experiments. PPN, integrating multilayer perceptron (MLP) and long short-term memory (LSTM) models, was developed using 135 DOE-based experiments on injection-molded polyamide-6 composite parts, compounded with aluminum diethylphosphinate and chopped carbon fibers, considering injection pressure, injection speed, packing time and packing pressure. PPN demonstrated superior predictive accuracy with a mean absolute percentage error (MAPE) of 0. 0557, outperforming previous studies employing single model approaches (0. 281 for MLP and 0. 147 for LSTM). PPN also achieved a low root mean squared error (RMSE) of 1. 1904 and high coefficients of determination (R2 > 0. 832) for predicted tensile properties with a prediction uncertainty under 2. 02 % quantified by a deep ensemble technique. Notably, PPN not only enhanced prediction accuracy by up to 81. 8 % compared to the conventional DOE, but also successfully predicted tensile properties for 1000 virtual experiments, providing results superior to the DOE analysis. This study demonstrated that PPN can effectively predict tensile properties for unseen process parameters, showing its potential for rapid design optimization by enabling comprehensive virtual experiments.

Details DOI

EAAI Journal 2025 Journal Article

Surface defect detection using distributed features

The Van Le
Jordan Daniel Joshua
Taehwan Kim
Jinhyuk Lee
Seong Han Kim
Jin Young Lee

Surface defect detection is crucial in manufacturing and has attracted increasing attention due to its significant impact on product quality and operational efficiency. With advancements in artificial intelligence (AI), deep learning based networks have become widely utilized for automated surface defect detection. However, most existing networks rely on a standard convolutional layer (Conv), which mainly extracts local information due to their limited learning region defined by kernel size. To address this limitation, we propose a surface defect detection network utilizing distributed features (DF-SDD), which consists of a distributed feature extraction block (DFEB) for feature extraction and a spatial attention gate (SAG) for feature fusion. DFEB effectively captures local and long-range information by introducing a distributed convolutional layer (DConv) and a distributed depthwise convolutional layer (DDWConv), which offer a broader learning area while maintaining low complexity. SAG merges features from different levels to enhance spatial information across multiple scales. Experimental results show that DF-SDD outperforms state-of-the-art networks on surface defect datasets in terms of both performance and complexity.

Details DOI

NeurIPS Conference 2010 Conference Paper

Sparse Coding for Learning Interpretable Spatio-Temporal Primitives

Taehwan Kim
Gregory Shakhnarovich
Raquel Urtasun

Sparse coding has recently become a popular approach in computer vision to learn dictionaries of natural images. In this paper we extend sparse coding to learn interpretable spatio-temporal primitives of human motion. We cast the problem of learning spatio-temporal primitives as a tensor factorization problem and introduce constraints to learn interpretable primitives. In particular, we use group norms over those tensors, diagonal constraints on the activations as well as smoothness constraints that are inherent to human motion. We demonstrate the effectiveness of our approach to learn interpretable representations of human motion from motion capture data, and show that our approach outperforms recently developed matching pursuit and sparse coding algorithms.

PDF Details