Deriving optimal atomic layer deposition process conditions using machine learning

The increasing complexity and high aspect ratios of next-generation semiconductor structures have intensified pattern loading effects in atomic layer deposition (ALD) processes. These effects result in non-uniform thin-film deposition rates and thickness variations. Deriving optimal process conditions to ensure consistent thin-film deposition is essential for maintaining substrate uniformity and enhancing device performance. Consequently, computational fluid dynamics (CFD) simulations have established themselves as effective tools for deriving process conditions. However, their high computational resource demands and inefficiency in adapting to changing conditions highlight inherent limitations in their application. To address these challenges, this study proposes the atomic layer deposition-gaussian process regression (ALD-GPR) model, integrating multi-layer perceptron (MLP) and gaussian process regression (GPR). The ALD-GPR model accurately predicts partial pressure, a key indicator of thin-film uniformity, achieving an RMSE of 0.0074 and approximately 18 times faster computation speed than CFD simulators, demonstrating its potential as an efficient alternative. Additionally, variance-based and difference-based metrics were developed to quantitatively evaluate uniformity and derive optimal conditions for achieving uniform thin films. These metrics provide a practical framework for assessing and enhancing process uniformity in ALD operations. The proposed ALD-GPR model and metrics optimize ALD processes and significantly improve computational efficiency and accuracy, providing an approach applicable to semiconductor manufacturing and other industries.