Multi-Scale Simulation of GaN Alkaline Wet Etching Leveraging a Machine Learning Potential

Gallium Nitride (GaN) is a versatile semiconductor with various advantages of wide bandgap, high electron mobility, high breakdown voltage, and exceptional optical properties. Wet etching of GaN in strong alkaline solutions is a crucial step in device fabrication such as micro-LEDs and high electron mobility transistors (HEMTs). This process involves intricate chemical interactions at the interface between GaN and the etchant solution. Although extensive experimental works have studied the etching process, a comprehensive atomistic understanding of the underlying mechanisms is still lacking because of high complexity of the system and difficulty in measurement. Ab initio molecular dynamics (AIMD) simulations have been widely employed to investigate reaction pathways at an atomic scale. However, they have significant computational costs, limiting the simulation size and time. Therefore, it would be difficult to gain insights into wet etching that involves diverse sluggish reactions and have significant size effects. In light of these challenges, recent advancements in machine learning potential (MLP) have provided promising alternatives. An MLP-based MD, such as those using a Behler-Parrinello type neural network potential (BPNNP), offers a more efficient approach for simulating complex chemical systems on a longer time scale.

In this work, we develop a multiscale computational model capable of simulating GaN wet etching in an aqueous potassium hydroxide (KOH) solution with high precision. We first construct an accurate and reliable BPNNP for simulating etching processes. To this end, we generate a training set by sampling AIMD trajectories of GaN surface, KOH aqueous solution, and their interface models at wide ranges of temperature and pressure encompassing supercritical phase. The model achieves an energy root-mean-square error (RMSE) of 20 meV/atom and a force RMSE of 0.2 eV/Å, accurately reproducing bulk GaN and solution properties as well as various reaction energies at the interface. The NNP-MD simulations are performed to investigate the etching behavior of polar and nonpolar surfaces of GaN in KOH solution by applying high-temperature conditions to accelerate chemical reactions. Our results show that the 1D GaN structure evolves from a slanted (r-surface) to a more flattened configuration (m-surface) during wet etching, which aligns with experimental observations. In addition, our calculations well describe the relative etching rates among various facets in experiments.

To bridge the time-scale gap between atomistic-scale simulations and macroscopic device behaviors, we utilize kinetic Monte Carlo (kMC) simulations. Our kMC model incorporates coarse-graining of atomic environments, allowing us to simulate large-scale morphological changes over extended time scales. We utilize on-the-fly enhanced sampling (OPES) to obtain free energy surfaces based on the developed BPNNP for chemical reactions that are likely to occur during wet etching. From the KMC simulations, we elucidate the selectivity between nonpolar surfaces of m-surface and a-surface and predict the etch rates at nm/min scale.

Our multiscale modeling framework, which combines NNP-based atomistic simulations, and kMC provides a comprehensive and detailed understanding of GaN alkaline wet etching. By elucidating the etching mechanisms, surface transformations, and the influence of etching parameters on kinetics and morphology, our findings contribute to the development of more precise and optimized processes in GaN-based device manufacturing. This integrated approach not only enhances control over the etching process but also enables more reliable predictions of device performance, ultimately advancing the efficiency and scalability of semiconductor production.