The global demand for computing power continues to surge with the advancement of artificial intelligence (AI), necessitating research into highly sophisticated semiconductor devices.
This trend emphasizes the need for advanced semiconductor manufactu...
The global demand for computing power continues to surge with the advancement of artificial intelligence (AI), necessitating research into highly sophisticated semiconductor devices.
This trend emphasizes the need for advanced semiconductor manufacturing processes capable of precise fabrication.
Atomic-scale processes such as atomic layer deposition (ALD) and atomic layer etch (ALE) are key technologies in achieving this level of miniaturization.
Simulation tools are vital in semiconductor process research, covering scales from the wafer down to the atomic level.
Among these, atomistic simulation is crucial for providing fundamental understanding and critical process design insights by interpreting atomic interactions based on physical laws at the smallest scale.
Traditionally, there is a trade-off where higher accuracy in describing atomic interactions requires significantly greater computational cost.
Machine learning interatomic potentials (MLIPs), a rapidly developing area, offer a breakthrough in this cost-accuracy relationship.
This thesis discusses the versatility of atomistic modeling integrated with MLIPs across various semiconductor processes, specifically focusing on ALD, plasma etching, and cryogenic etching.
Titanium nitride (TiN) is widely used in semiconductor devices as a diffusion barrier and metal gate.
To achieve precise film control, ALD using TiCl$_4$ and NH$_3$ is often implemented.
We developed a kinetic Monte Carlo (kMC) model to simulate this process, which successfully showed results consistent with experimental data regarding growth rate and Cl contamination.
Furthermore, we addressed the computational bottleneck in the kMC reaction finding step by exploring two acceleration methods: reaction searching automation using saddle point search (SPS) and speed enhancement via MLIPs.
SPS successfully identified relevant events, including gas evolution, used in conventional kMC models.
We also demonstrated the creation of an SPS-capable MLIP without prior knowledge on the system, showing that the combination of MLIP and SPS enables diverse event sampling in larger systems.
Plasma etching is a cornerstone of semiconductor manufacturing, involving complex physical phenomena from the interaction of various species (ions, neutrals, and radicals).
Focusing on ion beam etching as an atomistic representation of plasma etching, we used MLIPs to simulate the process and validate the results against experimental references.
For the well-established hydrofluorocarbon ion etching system of SiO$_2$ and Si$_3$N$_4$, we included random structures reflecting the various etching regime compositions and densities in the MLIP's primary training set.
This demonstrated a systematic way to create an MLIP for etching with minimal prior system knowledge.
The MLIP-driven simulations were validated against experimental results for etch yield, surface height, and surface composition.
We also performed an atomistic observation and mechanistic analysis of the transition from etching to deposition under specific conditions.
Cryogenic etching is crucial for processes with high etch rates like channel hole etching, offering superior performance by operating at low temperatures.
Optimal process design requires a deep understanding of gas-gas interactions under cryogenic conditions, alongside temperature control.
We used a recently developed pretrained universal MLIP (U-MLIP) to verify its capability in describing atomistic interaction, adsorption energy, and diffusion coefficients related to gas chemistry.
Focusing on the impact of gas transport via surface diffusion on etch rate at the etch profile's sidewall, we modeled the situation where HF (etchant) and IF$_5$ (additive gas) are present on ammonium fluorosilicate (AFS) under cryogenic conditions.
The U-MLIP showed good performance for adsorption energy and diffusion coefficient, and its accuracy was further enhanced after fine-tuning.
We observed an increase in the diffusion coefficient with the presence of the additive gas, which we explained in terms of HF chain length.
Additionally, we verified the validity of the assumed etchant and additive gas amounts using a multi-layer adsorption model.
Collectively, these individual studies confirm the significance of atomistic modeling and MLIPs in providing valuable insights for key semiconductor fabrication processes, including deposition and etching.
This work contributes to the fundamental understanding and advanced design of semiconductor manufacturing.
We believe the methodologies and findings presented here will pave the way for future developments in the field, further bridging the gap between computational simulation and experimental process engineering.