Industrial multi-machine data aggregation, AI-ready data preparation, and machine learning for virtual metrology in semiconductor wafer and slider production

Smart Manufacturing (SM), which is short for “Smart (Predictive, Preventive, Proactive) zero incident, zero emissions Manufacturing,” describes manufacturing’s digital transformation in which factories, supply chains and ecosystems are integrated, interoperable, and interconnected. Smart Manufacturing is rooted in AI, Machine Learned (ML), and Data Synchronized (DS) modeling to tap into invaluable operating data. By making data actionable at larger scales, SM opens new ways to increase productivity, precision, and process performance. Smart Manufacturing applied to front-end wafer manufacturing in the semiconductor industry offers significant opportunity to increase production throughput and ensure precision by increasing staff and operational productivity. Front-end wafer manufacturing involves multi tool operations for complex material processing that requires a high degree of precision and extensive product qualification. There is a high degree of commonality with semiconductor manufacturing tools, for example etching, that are well instrumented. Companies are already collecting large amounts of operational data from these tools that can be aggregated and leveraged for virtual metrology and other control, diagnostic, and management solutions. AI/ML/DS modeling involves monitoring the state of an operation in real-time to continuously learn and improve on human centered, automated, and autonomous actions. This operational data are embedded in invaluable machine, process, product, and material behaviors as interaction complexities, linearities/non-linearities, and dimensional effects. Because of machine commonalities, data can be selected to draw out operational value across machines. Today’s data science offers considerable capability for qualifying, assessing alignment and contribution, aggregating, and engineering data for more robust modeling. We refer to this as a Data-first strategy to process, engineer and model with AI-Ready data. In this paper, we address AI-Ready data for a virtual metrology solution focused on etching measurement PASS/FAIL classification and milling depth prediction regression tasks using operational data from production machine tools. If the quality of the product can be predicted, the productivity of the metrology process can be increased, which in turn increases the productivity of the overall operation. In a previous paper, we considered how to aggregate data from different etch tools in the same processes at different factories within Seagate Technology and proposed a method for data aggregation and demonstrated its value (Ou et al., 2024). The present paper considers how to process and engineer datasets from two different etch tool processes: wafer and slider production. The data processing approaches when used systematically with appropriate ML algorithms demonstrate the potential for reducing metrological interventions in semiconductor manufacturing. Advanced machine learning techniques are used to tackle the modeling challenges of a low failure rate and limited operational variability. XGBoost, a gradient descent-based tree algorithm, outperforms the commonly used Feedforward Neural Networks (FNN) in terms of training speed and resource utilization for binary-classifications, as well the performance criterion in ROC-AUC score (classification), Median Absolute Error (regression) and $R^2$ value. Principal Component Analysis (PCA) effectively reduces the dimensionality of the data and overfitting, while retaining vital variances and significantly reducing noise. Data aggregation with separated scaling harmonizes inputs from diverse manufacturing tools and significantly improves the efficacy and versatility of combining multiple datasets to improve model performance. A live updating transfer learning approach, that periodically updates the FNN models in real-time using Stochastic Gradient Descent (SGD) with individual data points, addresses process drift, and markedly improves predictive accuracy. For the slider production tools, data augmentation with linear Mixup, overcomes a short recording period, enriches the training dataset, and significantly reduces error metrics.