Tackling the "CO puzzle": AI-driven approaches to provide a novel method in surface chemistry with speed and precision

Understanding how molecules interact with metal surfaces is fundamental to catalysis and surface chemistry. However, traditional computational methods face a trade-off: achieving high accuracy often involves prohibitively expensive calculations, limiting large-scale or complex studies. A research team from Peking University have used a machine learning framework to create advanced exchange-correlation functionals within density functional theory (DFT). This approach enables accurate predictions of CO adsorption energies and site preferences on Cu(111) and Rh(111) surfaces with computational costs comparable to standard methods. The innovative DeePKS framework not only reproduces hybrid functional accuracy but also exhibits transferability across different adsorbate coverages, opening promising avenues for catalyst discovery and surface reaction modelling.

Understanding the interaction between molecules and transition metal surfaces is fundamental to designing efficient catalysts for industrial applications. Carbon monoxide (CO) adsorption on metals like Copper (Cu) and Rhodium (Rh) serves as a critical benchmark for testing the accuracy of these simulations. However, standard Density Functional Theory (DFT) methods face a significant challenge known as the "CO adsorption puzzle." The widely-used approximation Perdew-Burke-Ernzerhof (PBE) consistently fails to replicate experimental observations. This functional often fails to predict the correct preferred adsorption sites of CO. While higher-level methods like hybrid functionals (e.g., HSE06) can improve accuracy, they demand significant computational resources, making routine analysis and large-system simulations impractical. Additionally, previous low-cost workarounds often lack transferability, failing to provide a unified solution optimized for multiple surfaces simultaneously. Therefore, developing a transferable framework that achieves hybrid-functional accuracy with high efficiency is essential for the broad exploration of catalytic materials.

Addressing this challenge, the researchers employed the Deep Kohn-Sham (DeePKS) framework to train machine-learned exchange-correlation functionals that enhances the exchange-correlation functional used in DFT calculations. By training neural network models on high-accuracy data, the researchers successfully developed a universal, transferable functional capable of predicting adsorption energies, site preferences, and potential energy surfaces with near-hybrid functional accuracy. They developed both system-specific models for Copper (Cu) and Rhodium (Rh) (111) surfaces, and more notably, a unified model capable of seamlessly handling both metals in a single framework. The models accurately recover the experimental preference for top-site adsorption which approximations like PBE persistently fail to predict, thus providing a new possibility to resolve the "CO adsorption puzzle". The unified model achieves accuracy comparable to the single-metal models, yielding energy differences of only about 10 meV compared to expensive hybrid functional benchmarks. In addition, the models accurately reproduce the full potential energy surfaces and relaxed atomic structures, rather than just single energy points. The approach also demonstrates satisfactory generalization, correctly predicting adsorption behaviors at various unseen adsorbate coverages that were not included in the training set. These results highlight a promising path toward universal models that enable the exploration of complex catalytic systems with hybrid-functional accuracy at a substantially reduced cost.

The Future: Future research will focus on capturing the complete electronic structure of materials and improving the model's universal transferability.

While the current work provides a new AI solution to resolve the energy puzzle on metal surfaces, the research team plans to leverage the unique advantage of the DeePKS framework: its ability to stay within the quantum mechanical theory. Unlike conventional AI force fields that only predict total energy and atomic forces, future work will explicitly incorporate electronic properties, such as energy levels, directly into the training process. This will allow the model to predict electronic structure with high accuracy, offering deeper insights into catalytic mechanisms that other methods cannot provide. Additionally, the researchers aim to introduce fundamental physical constraints and exact conditions into the model. This improvement is expected to further enhance the model's ability to transfer its knowledge across diverse material systems, significantly reducing the amount of training data required for new discoveries.

The Impact: This work offers a promising pathway to simulating complex surface reactions with hybrid-functional accuracy and opens new avenues for the efficient, low-cost design of advanced catalysts for industrial applications.

The research has been recently published in the online edition of AI for Science, a diamond open access journal co-published by IOP Publishing and Songshan Lake Materials Laboratory that focuses on the transformative application of artificial intelligence in driving scientific innovation.

Reference: Xinyuan Liang, Renxi Liu and Mohan Chen*. Investigating CO Adsorption on Cu(111) and Rh(111) Surfaces Using Machine Learning Exchange-Correlation Functionals[J]. AI for Science, 2025, 1(2): 025007. DOI: 10.1088/3050-287X/ae21fa

---

---