Predicting Adsorption Energies Using Multifidelity Data.

Area of Science:

• Materials Science
• Computational Chemistry
• Surface Science

Background:

• Accurately predicting adsorbate binding energies on transition-metal surfaces is crucial for catalyst design.
• Existing methods often rely on high-fidelity data, which is computationally expensive or experimentally challenging to obtain.
• Integrating diverse data sources can potentially improve model accuracy and efficiency.

Purpose of the Study:

• To introduce and evaluate multitask Gaussian processes (MT-GPs) for modeling adsorbate binding energies.
• To demonstrate the effectiveness of MT-GPs in leveraging both low-fidelity and high-fidelity data.
• To compare the performance of MT-GPs against single-task models.

Main Methods:

• Utilized multitask Gaussian processes (MT-GPs) to model binding energies.
• Integrated data from multiple sources, including density functional theory (DFT) and random phase approximation (RPA) computations, as well as experimental data.
• Conducted two case studies: one with purely computational data and another with combined computational and experimental data.

Main Results:

• MT-GPs achieved high-fidelity modeling of binding energies by effectively combining data from various sources.
• In both case studies, MT-GPs significantly outperformed single-task models trained on individual data sources.
• The method demonstrated superior performance when fusing datasets.

Conclusions:

• MT-GPs offer a powerful approach for learning improved models from fused datasets of varying fidelity.
• This method maximizes model accuracy within constrained computational and experimental budgets.
• The findings suggest MT-GPs are a valuable tool for materials modeling and discovery.