When correlation becomes causation: deep learning for complex anomaly detection in semiconductor yield

We present an explainable anomaly detection framework for semiconductor photomask manufacturing data, designed to identify and interpret complex abnormal patterns in multivariate time-series data. The dataset was constructed by aggregating high-frequency sensor measurements from photomask production processes. To improve robustness under industrial noise, we applied appropriate normalization techniques and implemented a dynamic thresholding mechanism for anomaly score segmentation. While this forms the foundation of our anomaly detection pipeline based on attention-driven reconstruction, the main contribution of this work lies in the explainability layer developed as an extension of the detection output. After identifying candidate anomalies using a transformer-based model, we extract feature-wise reconstruction errors and attention-weighted deviations to localize affected segments. These segments are formatted into structured prompts and passed to large language models (LLMs), which are guided to generate human-readable explanations. Through prompt engineering, the LLM compares each anomalous segment with its surrounding baseline context, summarizes the propagation of anomalies across correlated features, and provides root cause reasoning aligned with domain knowledge. These explanations not only clarify why the anomaly was detected, but also identify the most responsible variables and outline potential implications for manufacturing. To the best of our knowledge, this is the first work to employ LLMs for interpretable explanations of multivariate time-series anomalies, rather than detection alone. Experiments on proprietary semiconductor datasets show that our backbone anomaly detector achieves F1-score ∼ 0.98 under industrial noise, while LLM-based explanations reach type accuracy of 0.625 with faithfulness up to 1.0 on our evaluation. This demonstrates the feasibility of combining deep learning detection with reasoning-oriented language models, enabling transparent interpretation of subtle faults in yield-critical semiconductor manufacturing.