(314g) Fault Detection for an Atomic Layer Etching Process Using Machine Learning

Oftentimes, semiconductor production faces obstacles that are attributed to the stringent quality specifications as a consequence of the continued downscaling and intricate designs for transistors. A vital procedure in the fabrication of transistors is the use of sequential cycles of atomic layer deposition (ALD) and atomic layer etching (ALE) to deposit of etch monolayer of high-κ oxide films to minimize the observance of short-channel effects and preserve the integrity of the transistor performance [1]. To facilitate the process of designing and optimizing atomic layer processes for industrial applications, in silico modeling, in the form of multiscale computational fluid dynamics (CFD), has paved the way for generating realistic data that is synonymous to industrial data. However, there lacks an effective way of replicating the stochastic frequency of disturbances that lead to product nonconformance, deformities, and failure [2]. Thus, synthetic data requires a cross-validation procedure with industrially relevant data to further improve the accuracy and reproducibility of modeling data.

In industrial practice, these thin-layer deposition and etching processes observe unpredictable defects that are consequential to product failure and are difficult to identify due to the numerous input parameters supplied [3]. To overcome this burden, machine learning, as a pivotal tool for establishing models for systems with numerous dimensions, is applied to train and test industrial data for an ALE process. Feedforward neural networks (FNNs) are employed to train industrial data obtained from Seagate Technologies comprising subsets of input conditions that are classified by reactor or tool type, substrate material, and process operating conditions, and a single, binary output parameter that is defined by a pass or fail classifier. These FNNs first utilize encoder tools for textual variables, which are then trained with single-tool and aggregated-tool data to test their prediction of fault detection across other tool types [4]. Through this manner, a prediction on the product defect rate can be made prior to development of a reactor; thereby enabling the development of more reliable processes. Additionally, the relevancy of this industrial data, particularly the defect ratio, will be further applied to prior in silico multiscale models for an ALE process [5] to enhance the accuracy of prior simulated models by introducing stochastic disturbances that are uniformly distributed across an average product defect rate obtained from the industrial data.

• Jurczak, M., Collaert, N., Veloso, A., Hoffmann, T., Biesemans, S., 2009. Review of FINFET technology. In: 2009 IEEE International SOI Conference, 1–4, Foster City, CA, USA.
• Maroudas, D., 2000. Multiscale modeling of hard materials: Challenges and opportunities for chemical engineering. AIChE Journal, 46, 878–882.
• Zhang, C., Yella, J., Huang, Y., Qian, X., Petrov, S., Rzhetsky, A., Bom, S., 2021. Soft sensing transformer: Hundreds of sensors are worth a single word. In: 2021 IEEE International Conference on Big Data (Big Data), 1999–2008, Orlando, FL, USA.
• Iskandar, J., Moyne, J., Subrahmanyam, K., Hawkins, P., Armacost, M., 2015. Predictive Maintenance in semiconductor manufacturing. In: 2015 26th Annual SEMI Advanced Semiconductor Manufacturing Conference (ASMC), 384–389, Saratoga Springs, NY, USA.
• Yun, S., Tom, M., Ou, F., Orkoulas, G., Christofides, P. D., 2022. Multiscale computational fluid dynamics modeling of thermal atomic layer etching: Application to chamber configuration design. Computers & Chemical Engineering, 161, 107757.