(375u) Time-Series Multiscale Computational Fluid Dynamics Data Modeling with Transformers for Atomic Layer Processing

Atomic layer processing is a crucial procedure in which thin film coatings are deposited or etched onto transistor surfaces to reduce short-channel effects and maintain robust properties for finished semiconductor materials. Through this manner, monolayers of high-κ oxide films are deposited or etched in sequential cycles in a self-limiting behavior. However, maintaining this self-limiting tendency is a challenging task due to the lack of available data for process optimization and the difficulties attributed to process scale up from experimental scales. With the integration of micro-electromechanical systems (MEMS), on-line data can be extracted in discrete time steps by sensors to generate time-series data that is effective at observing the temporal behavior of the process in real-time [1]. An advantageous feature of time-series data is the ability to predict operating conditions of the reactor for subsequent times [2], which is applicable to the integration of online feedback control and model predictive control for these atomic layer processes. However, there are limitations in obtaining time-series data from experimental contexts, especially for processes that have not been integrated into industrial practice. Thus, in silico modeling is a beneficial route toward the generation of meaningful data that is reflective of processes in industry.

For this work, a two-dimensional multiscale computational fluid dynamics (CFD) model that establishes a codependent framework between mesoscopic kinetic Monte Carlo (kMC) and macroscopic CFD simulations is employed to produce time-series data for a range of operating conditions (e.g., precursor flow rates) and reactor configurations (e.g., gap distances). This multiscale CFD model will integrate a previously developed thermal atomic layer etching process for the etching of Al₂O₃ films [3] to extract meaningful spatiotemporal data at various simulation settings. However, multiscale CFD modeling is a time-consuming task that requires an abundance of computational resources to produce data efficiently. Thus, a predictive model trained on this spatiotemporal data is vital to pursuing operational decision-making tasks for reactor configurations that have not yet been established. To facilitate the process of building a predictive model, a transformer is employed due to its outperforming of state-of-the-art methods in natural language processing [4]. Lastly, a comparison of various aggregated- and singular-tool models are also conducted to determine the accuracy of the prediction made by the transformer models.

[1] Cheng, C. L., Chang, H. C., Chang, C. I., Fang, W., 2015. Development of a CMOS MEMS pressure sensor with a mechanical force-displacement transduction structure. Journal of Micromechanics and Microengineering, 25, 125024.

[2] Ahn, J., Kim, H.-Y., Cho, S.-H., Kim, H.-J., Kim, H., Choi, H., Ham, D., 2023. Semiconductor equipment health monitoring with multi-view data, In: Winter Simulation Conference (WSC), 2322–2332, San Antonio, TX, USA.

[3] 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.

[4] 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.