(599b) Pre-Trained Universal Catalyst Nanoparticle Model for Screening Catalytic Activity in General Alloy Crystals

Heterogenous catalysts play a critical role in many industrial processes as they can bring both environmental and economic advantages by enabling processes to happen, reducing energy consumption, and increasing production. The Discovery of novel catalysts or non-previous alternatives with high performance is, therefore, an important task. However, conventional experiment-based catalyst discovery and design are time-intensive and costly. This approach may easily take years or even decades from the initial idea formation to the final industrial implementation. Repeated experiments would be required to synthesize and test catalysts with different active components, synthesis strategies, and process parameters. For practical considerations, we would use our knowledge and understanding of the target reaction to focus on a few hundred selected catalysts of the entire catalyst space. Given that the actual catalyst space is normally huge, it may be possible to find other catalyst candidates with higher performance or lower cost if we expand our search space. However, this would require a much cheaper way to evaluate the performance of catalysts. Recent advances in computational and machine learning methods can potentially expand this search range. For a given catalytic reaction with a well-studied pathway, computational methods can avoid experiments and estimate catalyst activity via specific microkinetic descriptors in hours; and machine learning surrogates can further accelerate the estimation within seconds.

While most of the current related works focus on a specific reaction and predict the activity of crystals based on the activity of individual stable facets [1-4], we notice two major limitations of the current methods. First, most common catalyst synthesis methods, such as impregnation and precipitation, result in crystal nanoparticles of Wulff shapes enclosed by different exposed facets. Moreover, some relatively unstable facets with small exposed areas may also contribute to high activity. Therefore, the activity of individual stable facets may not reflect the activity of the actual catalyst nanoparticle. Second, considering the similarity of DFT energy calculations, a universal surrogate could be developed from a large DFT dataset to avoid data generation from zero for every new catalyst discovery task. Therefore, we aim to develop a pre-trained universal catalyst nanoparticle model. The model would predict the overall activity of each catalyst based on the activity of the exposed facets of its nanoparticle. The developed model can provide a fast screening of the activity of overall catalyst nanoparticles in a large catalyst space for any heterogeneous reactions with known pathways or descriptors.

In this work, we propose a novel framework for a catalyst nanoparticle model. A catalyst space is formulated by retrieving the crystal structures of all thermodynamically stable alloy crystals from the available databases, such as Material Project. To accurately estimate the activity of a specific crystal in the catalyst space, we will develop machine learning models to predict its crystal nanoparticle’s shape, facet exposure, and activity of the exposed facet. The overall activity of the crystal would then be determined from all exposed facets. It should be noted that related catalyst property databases are required to train the models mentioned above. This work will leverage existing databases and also generate additionally required databases for model development.

The developed catalyst nanoparticle model can be directly applied to give initial recommendations for catalyst discovery tasks and fine-tuned with additional data of a specific reaction to give more accurate predictions and narrowed recommendations for experimental validation.

[1] D. Behrendt, S. Banerjee, C. Clark, and A. M. Rappe, “High-Throughput Computational Screening of Bioinspired Dual-Atom Alloys for CO2 Activation,” Journal of the American Chemical Society, vol. 145, no. 8, pp. 4730-4735, 2023/03/01 2023, doi: 10.1021/jacs.2c13253.

[2] H. A. Doan, C. Li, L. Ward, M. Zhou, L. A. Curtiss, and R. S. Assary, “Accelerating the evaluation of crucial descriptors for catalyst screening via message passing neural network,” Digital Discovery, 2023.

[3] K. Tran and Z. W. Ulissi, “Active learning across intermetallics to guide discovery of electrocatalysts for CO2 reduction and H2 evolution,” Nature Catalysis, vol. 1, no. 9, pp. 696-703, 2018/09/01 2018, doi: 10.1038/s41929-018-0142-1.

[4] S. Back, J. Yoon, N. Tian, W. Zhong, K. Tran, and Z. W. Ulissi, “Convolutional Neural Network of Atomic Surface Structures To Predict Binding Energies for High-Throughput Screening of Catalysts,” The Journal of Physical Chemistry Letters, vol. 10, no. 15, pp. 4401-4408, 2019/08/01 2019, doi: 10.1021/acs.jpclett.9b01428.