Physics-structured graph neural networks for interpretable fault detection in semiconductor manufacturing processes

Smart manufacturing has transformed semiconductor fabrication into a highly data-driven environment where large numbers of in-situ sensors continuously generate process monitoring data across multiple production tools. Leveraging these data streams to support yield improvement and fault detection has become an important challenge, particularly because conventional physical metrology is often costly, time-consuming, or destructive. Machine learning-based soft sensing has therefore emerged as a practical alternative for predicting process outcomes using routine sensor measurements and contextual process information. Previous studies have demonstrated that classical machine learning models such as gradient boosting decision trees (GBDTs) and feedforward neural networks (FNN) can improve pass/fail classification performance when applied to individual tools or aggregated intra-tool data (i.e., tools of the same class). In this work, we extend these approaches to a broader range of semiconductor manufacturing processes (i.e., data from tools belonging in different classes as well data from the same tool when making different products), and propose a physics-structured graph neural network (GNN) framework for PASS/FAIL prediction. In addition, modern tabular deep-learning models, including TabM and TabR, are evaluated as complementary benchmarks to determine whether strong neural tabular learning can improve fault-detection performance beyond conventional GBDT models and physics-structured graph learning. Instead of modeling individual sensors directly, sensor measurements are organized into physically meaningful process subsystems, each representing a group of related sensors describing process dynamics. These subsystem representations are combined with categorical contextual variables and processed through a graph attention architecture to capture interactions among process modules. TabM is used as a parameter-efficient ensemble-style neural tabular model, while TabR introduces retrieval-augmented prediction by comparing each production run with similar historical training runs. The models are evaluated using production data collected over an 18-month period. For each tool, the most recent 20% of the timeline is reserved for testing while the earlier data are used for training and validation. To ensure statistical robustness, only tools with more than 8000 observations within the study period are included. Results demonstrate that the proposed physics-structured GNN improves predictive performance on several tools compared with conventional machine learning approaches while also providing interpretable attention weights that highlight the relative contribution of different process subsystems to predicted failures. The TabM and TabR results further demonstrate that modern tabular deep-learning models can be useful complementary predictors, especially on tools with richer failure information, although they do not consistently replace the optimized LightGBM benchmark under severe class imbalance. These results demonstrate the potential of physics-guided graph learning and complementary tabular deep-learning methods for scalable and interpretable fault detection in semiconductor manufacturing environments.