Investigating oxygen reduction at gold-water interface with machine learning potentials

Fuel cell devices are considered as an ideal solution for the transition of a sustainable future, with their performance significantly influenced by catalysts that reduce the overpotential of the oxygen reduction reaction (ORR). A prerequisite for optimizing ORR catalysts is an in-depth understanding towards the reaction mechanisms in an atomistic level, which is often achieved by density functional theory (DFT) calculations. However, the expensive computational cost of DFT has significantly limited its length-scale. Machine-learned interatomic potentials (MLIPs) have emerged as powerful tools in the domain of atomistic simulations due to their exceptional computational efficiency and ab-initio level accuracy. At the heart of creating superior MLIPs for specific applications lies the imperative for high-quality data. Yet, acquiring high-quality data for vast chemical spaces remains challenging, often requiring costly ab-initio simulations. This thesis focuses on creating a robust framework to accelerate the generation of MLIPs and leverage it to gain insights into ORR mechanisms on gold surfaces.

Firstly, we designed an autonomous active learning workflow CURATOR for training highfidelity graph neural network potentials for atomistic simulations. With the well-designed batch active learning algorithms, it can efficiently acquires high-quality data for optimizing model improvement during retraining. By integrating advanced neural networks with reliable uncertainty quantification techniques, CURATOR ensures accurate and efficient data acquisition, reducing human efforts and computational costs for MLIP construction. Additionally, it includes trustworthy and efficient uncertainty estimation techniques. By integrating different key components, this workflow is able to autonomously manage the complex tasks for generating MLIPs.

Subsequently, we investigated the ORR at confined Au(100)-water interface by using MLIPs-accelerated metadynamics. Combining MLIPs with enhanced sampling techniques allowed our simulations to achieve time-scales beyond the reach of conventional DFT. This framework vividly showcased the full ORR reaction process, pinpointing an associative ORR mechanism on Au(100) with a low reaction barrier, aligning with experimental results. This framework shed the light on modeling complex chemical reactions under complex ambient conditions.

Having verified the predictive power of the developed framework, we extended our research systems to other primary facets of gold, using a larger simulation box to better capture the reaction dynamics. Our simulations simulations revealed the notable presence of *H₂O₂ on Au(110) and Au(111). This observation is in good agreement with experimental results and shed the light on optimizing the performance of gold-based ORR catalysts.