Challenges and Opportunities of Pretrained Machine Learning Interatomic Potentials in Heterogeneous Catalysis

Area of Science:

• Materials Science
• Computational Chemistry
• Catalysis

Background:

• Accurate modeling of surface reactivity is crucial for catalyst design.
• Density functional theory (DFT) is the primary computational method for atomistic understanding.
• DFT calculations are computationally expensive.

Purpose of the Study:

• To provide an overview of state-of-the-art machine learning interatomic potentials (MLIPs) for heterogeneous catalysis.
• To assess MLIPs as "out-of-the-box" tools for catalysis research.
• To discuss the potential of MLIPs to democratize computational catalysis.

Main Methods:

• Summarize different families of MLIPs and their training processes.
• Apply pretrained MLIP models to heterogeneous catalysis problems.
• Critically evaluate model transferability and integration challenges.

Main Results:

• MLIPs show potential to match DFT accuracy with significantly reduced computational cost.
• Pretrained MLIPs can be applied to heterogeneous catalysis problems.
• Challenges remain in model transferability, standardization, and achieving reliable predictive power.

Conclusions:

• MLIPs represent a significant advancement in computational catalysis.
• Standardized protocols are needed to benchmark MLIP performance.
• Further development is required to overcome hurdles for widespread, reliable use of MLIPs.