Multimodal Anomaly Detection for Semiconductor Manufacturing Material Quality

In the semiconductor industry, as processes continue to miniaturize, monitoring material quality has become increasingly critical. Traditional methods, which largely rely on statistical analysis and expert-driven heuristics, are time-consuming and costly. This study introduces an approach using a multimodal structure combined with autoencoders to incerease the efficiency of the monitoring process, leading to reductions in both time and cost. The proposed method uses both time series and image data, which are transformed from raw semiconductor process material data through the Markov transition field technique. By integrating masking into the input data, processed separately by distinct autoencoders, the system effectively extracts and concatenates features. These are subsequently passed through a self-attention layer and reconstructed by each decoder to train the model. Experimental results demonstrate that the proposed multimodal approach significantly enhances anomaly detection capabilities, outpeforming traditional methods. The findings suggest substantial improvements in the efficiency and accuracy of quality monitoring in semiconductor manufacturing processes.