Presentation Information
[15p-70A_101-1]Computation of Al2O3 ALD by trimethylaluminum with Kinetic Monte Carlo and neural network potential
〇(M2)Yichen ZOU1, Yuxuan Wu1, Jun Yamaguchi1, Noboru Sato1, Atsuhiro Tsukune1, Yukihiro Shimogaki1 (1.The Univ. Tokyo)
Keywords:
ALD,KMC,NNP
Atomic layer deposition (ALD) is a key thin-film growth technique for advanced semiconductor and energy applications due to its atomic-scale thickness control and excellent conformality. Although the trimethylaluminum (TMA)–H2O process for Al2O3 is one of the most extensively studied ALD systems, quantitatively linking atomistic surface reactions to experimentally observed growth behavior remains challenging because of complex reaction networks, finite-temperature effects, and steric constraints. Computational modeling is therefore essential for establishing a mechanistic and predictive description of ALD kinetics.
In this study, we develop a kinetic Monte Carlo (KMC) framework for the TMA–H2O ALD process by integrating molecular dynamics (MD) simulations with a neural-network interatomic potential (NNP). The NNP, trained at near–density-functional-theory accuracy, enables efficient finite-temperature sampling of surface reactions and adsorption processes, providing thermodynamically consistent kinetic input for long-timescale KMC simulations. Reaction rate constants for surface reactions are evaluated from Gibbs free-energy barriers obtained via NNP-based sampling. Adsorption kinetics are described using a gas-kinetic formulation, in which the sticking probability of TMA is determined directly from MD collision simulations based on chemically irreversible adsorption events.
The resulting simulations quantitatively reproduce key features observed in in situ quartz-crystal microbalance measurements, including rapid mass uptake during the TMA pulse, saturation behavior with a mass gain of 36.8 ng/cm2, the negative mass excursion during the H2O pulse, and purge-induced desorption. These results demonstrate that the combined NNP–MD–KMC approach provides a physically grounded and predictive description of ALD growth kinetics.
In this study, we develop a kinetic Monte Carlo (KMC) framework for the TMA–H2O ALD process by integrating molecular dynamics (MD) simulations with a neural-network interatomic potential (NNP). The NNP, trained at near–density-functional-theory accuracy, enables efficient finite-temperature sampling of surface reactions and adsorption processes, providing thermodynamically consistent kinetic input for long-timescale KMC simulations. Reaction rate constants for surface reactions are evaluated from Gibbs free-energy barriers obtained via NNP-based sampling. Adsorption kinetics are described using a gas-kinetic formulation, in which the sticking probability of TMA is determined directly from MD collision simulations based on chemically irreversible adsorption events.
The resulting simulations quantitatively reproduce key features observed in in situ quartz-crystal microbalance measurements, including rapid mass uptake during the TMA pulse, saturation behavior with a mass gain of 36.8 ng/cm2, the negative mass excursion during the H2O pulse, and purge-induced desorption. These results demonstrate that the combined NNP–MD–KMC approach provides a physically grounded and predictive description of ALD growth kinetics.