The semiconductor industry grew from $555.9 billion in 2021 to $574.1 billion in 2022, making for a
3.3% increase [1]. This implies that, to meet the ever-growing demand, the global semiconductor
manufacturing capability must also increase by 3%, year over year. However, manufacturing operations
are often bottlenecked by product metrology steps that are necessary to ensure product quality. One
method of improving manufacturing efficiency is to use data-driven models to predict product quality,
allowing companies to drastically decrease, or even eliminate, these time-consuming product metrology
steps [3]. A previous collaboration with Seagate has shown that Neural Networks and data aggregation
techniques can be used in conjunction to achieve high predictive performance for critical etching process
layers [4]. However, other process steps, such as a stepper operation, have unique challenges such as
time-based data drift that require more specialized data modeling techniques to solve.
A common issue in manufacturing environments is the presence of time-varying process drifts that are
periodically reset after maintenance cycles. To combat this, a machine learning technique known as
Transfer Learning can be implemented. Transfer Learning freezes the structure of an initially developed
Neural Network and only allows its weights to be slightly changed, or updated, on a small subset of data
[2]. Another common issue is the lack of sufficient data, which can often be attributed to incomplete or
faulty measurements resulting from the degradation or malfunction of sensing equipment. Additionally,
limited data availability may arise when the equipment is deployed online for only a short duration,
thereby restricting the amount of data that can be used for training. Thus, data Mixup can be used to
augment small datasets [5]. By evaluating the performance of Neural Networks at predicting process
faults by itself, with Transfer Learning, and with data Mixup for a variety of process layers, the efficacy
of these techniques will be shown in a real, industrial settings. Additionally, by analyzing the process data
and which technique is best suited for it, generalizations regarding what kind of processes need what kind
of treatments can also be extracted. Thus, the results will describe both the most effective data processing
techniques and what sort of process data needs these techniques the most.
[1] Casanova, 2023. Chip sales rise in 2022, especially to auto, industrial, consumer markets.
[2] Jiang, Y., Yin, S., Dong, J., Kaynak, O., 2020. A review on soft sensors for monitoring, control,
and optimization of industrial processes. IEEE Sensors Journal 21, 12868–12881.
[3] Kadlec, P., Gabrys, B., Strandt, S., 2009. Data-driven soft sensors in the process industry.
Computers & Chemical Engineering 33, 795–814.
[4] Ou, F., Wang, H., Zhang, C., Tom, M., Bom, S., Davis, J.F., Christofides, P.D., 2024. Industrial
data-driven machine learning soft sensing for optimal operation of etching tools. Digital
Chemical Engineering 13, 100195
[5] Zhang, L., Deng, Z., Kawaguchi, K., Ghorbani, A., Zou, J., 2020. How does mixup help with
robustness and generalization? arXiv preprint arXiv:2010.04819
