npj Computational Materials

Machine-learning driven global optimization of surface adsorbate geometries

Hyunwook Jung, Lena Sauerland, Sina Stocker, Karsten Reuter & Johannes T. Margraf

npj Computational Materials 9: 114(2023)

doi.org/10.1038/s41524-023-01065-w

Published online: 26 June 2023
Abstract| Full Text | PDF OPEN

Abstract: The adsorption energies of molecular adsorbates on catalyst surfaces are key descriptors in computational catalysis research. For the relatively large reaction intermediates frequently encountered, e.g., in syngas conversion, a multitude of possible binding motifs leads to complex potential energy surfaces (PES), however. This implies that finding the optimal structure is a difficult global optimization problem, which leads to significant uncertainty about the stability of many intermediates. To tackle this issue, we present a global optimization protocol for surface adsorbate geometries which trains a surrogate machine learning potential on-the-fly. The approach is applicable to arbitrary surface models and adsorbates and minimizes both human intervention and the number of required DFT calculations by iteratively updating the training set with configurations explored by the algorithm. We demonstrate the efficiency of this approach for a diverse set of adsorbates on the Rh(111) and (211) surfaces.

摘要: 在计算催化研究中，分子吸附剂在催化剂表面上的吸附能是关键描述符。然而，对于经常遇到的相对较大的反应中间体，例如，在合成气转化中，多种可能的结合模式导致复杂的势能面（PES）。这意味着找到最优结构是一个困难的全局优化问题，从而导致许多中间体的稳定性存在显著的不确定性。为了解决这个问题，我们提出了一种用于表面吸附剂几何结构的全局优化协议，该协议实时训练代理机器学习势能。这一方法适用于任意表面模型和吸附剂，并通过迭代将算法探索的构型实时更新到训练集中，最小化人为干预和所需DFT计算的数量。我们证明了这种方法在Rh（111）和（211）表面上多种吸附剂的效率。

Editorial Summary

Dancing on catalyst surfaces: machine-learning driven global optimization protocol for adsorbate geometries

In computational catalysis research, the adsorption energy of adsorbates on the catalyst surface is a critical parameter for assessing catalyst activity. Particularly in complex catalytic reactions such as syngas conversion, which involve larger reaction intermediates, multiple possible binding modes result in complex potential energy surfaces (PES). Therefore, the global optimization of the adsorbate's geometry on the catalyst surface is of paramount importance. To address this challenge, this study proposes a machine learning-based universal global optimization protocol. A team led by Professor Johannes T. Margraf from the Fritz-Haber-Institut der Max-Planck-Gesellschaft, Germany, proposed a machine learning-driven universal global optimization protocol based on Gaussian Approximation Potentials (GAP). They employ an active learning workflow to efficiently search for the global optimal geometry of adsorbates. The protocol's efficiency is validated through tests on various important reaction intermediates in ethanol synthesis on Rh(111) and (211) surfaces. By using the Gaussian Approximation Potentials (GAP) surrogate model to explore the potential energy surface, the protocol requires only a small number of single-point density functional theory (DFT) calculations for model training while comprehensively exploring configuration space to identify a diverse set of promising candidate structures. Importantly, in minima hopping (MH) simulations, the training set is iteratively updated on-the-fly to achieve high data efficiency, and the model's hyperparameters are automatically determined using heuristic methods, minimizing the need for human intervention. This protocol provides a powerful tool for computational catalysis research.

催化剂表面上的舞蹈：机器学习全局优化吸附剂几何结构

计算催化研究中，吸附剂在催化剂表面的吸附能是评估催化剂活性的关键参数。特别是在复杂的催化反应中，如合成气转化，涉及到较大反应中间体，多种可能的结合模式导致了复杂的势能面（PES），因此全局优化吸附剂在催化剂表面上的几何结构至关重要。为解决这一挑战，本研究提出一种基于机器学习的通用全局优化协议。来自德国马克斯普朗克弗里茨哈伯研究所的Johannes T. Margraf教授研究团队，提出了一种基于高斯近似势（GAP）的机器学习驱动的通用全局优化协议。他们采用主动学习工作流程，高效寻找吸附剂的全局最优几何结构。他们通过对Rh(111)和(211)表面上乙醇合成的多种重要反应中间体的测试，验证了该协议的高效性。使用高斯近似势（GAP）代理模型来探索势能面，只需要少量的单点密度泛函理论（DFT）计算进行训练模型，就可以全面探索构型空间，得到多个有潜力的候选结构，从而提高计算效率。最要的是，在最小值跳跃（MH）模拟中，通过迭代实时更新训练集实现高数据效率，同时采用启发式方法自动确定模型的超参数实现最小化人工干预。该协议为计算催化研究提供了强大的工具。