6-11. Construction of high-throughput computational platform and prediction of adsorption energies using machine learning methods

Tao-Tao Shi¹，Zhi-Hui Liu¹，Zhao-Chen Cai²，Yue Chen²，Xiu-Feng Wu², Zhao-Xu Chen^1*

1. Institute of Theoretical and Computational Chemistry, Key Laboratory of Mesoscopic Chemistry of MOE, School of Chemistry and Chemical Engineering, Nanjing University；2. Nanjing University

Abstract: Catalysts are of vital importance in chemical industry. Finding catalysts of high activity and selectivity and low cost and less detrimental is the desired goal. The traditional trial-and-error approach is expensive, time-consuming and inefficient. With the advent of high-throughput concept and the development of computer technology and computational chemistry, theoretically screening catalysts of good performance is possible. Nørskov et al. found activation energies correlate the stability of reaction intermediates for a class of catalytic reactions, indicating that adsorption energies are closely related to the catalytic activity of a catalyst. At present, most catalyst screening uses adsorption energies as the descriptor.

To speed up the development and designing of good performance catalysts, we have established a high-throughput computational platform based on micro-service architecture. 10² concurrent calculations can be achieved by connecting external computing resources. The calculation process can be real-time monitored, and the computational results can be automatically extracted and stored.

Considering the importance of adsorption energies and using the platform, we systematically calculated the adsorption energies of 12 kinds of atoms adsorbed on the stable surfaces of 38 metals. We also computed the adsorption energies of CO on 9056 layered alloys. The reason for choosing such systems is that CO adsorption energies are believed to be the descriptor for screening catalysts used for CO₂ conversion into CH₃OH. Machine learning methods were used to analyze the large amount of data produced, and prediction models were established and satisfactory results were obtained.

Conclusions and prospects: We have constructed a high-throughput computational platform for massive computation, screening and prediction of catalytic materials based on descriptors. We studied adsorption of 12 single atoms on 38 metal surfaces and CO on 9056 layered alloys. We employed machine learning methods to establish prediction models. The present work will play an important role in future high-throughput computation and material screening.

Keywords: High-throughput computing; Machine learning; Adsorption energy

高通量计算平台的构建与吸附能的机器学习预测

史涛涛¹，刘志慧¹，蔡赵辰²，陈悦²，武秀峰²，陈兆旭^1*

1. 南京大学化学化工学院介观化学教育部重点实验室理论与计算化学研究所；2. 南京大学

摘要：催化剂在化工生产中起着至关重要的作用。寻找高活性、高选择性、低成本和低危害的催化剂一直是人们追求的目标。传统的“试错法”成本高、耗时长且效率低。高通量概念的出现、计算机技术及计算化学的发展使理论上筛选品优催化剂成为可能。Nørskov等发现活化能和反应中间体稳定性相关，预示吸附能和催化剂活性密切相关。目前，大多数催化剂筛选都使用吸附能作为描述因子。

为加快催化剂的研发，我们建立了基于微服务架构的高通量计算平台。通过连接外部计算资源，该平台可达到10²级并发计算，并实现计算过程的实时监控和计算结果自动提取及存储。

鉴于吸附能的重要性，我们利用该平台系统地计算了12种原子在38种金属的稳定表面上的吸附。由于CO吸附能被认为是筛选CO₂制CH₃OH的催化剂的描述因子，我们研究了CO在层状合金上的吸附。在此基础上，采用机器学习方法对获得的大量数据进行了分析，建立了预测模型，取得了良好的结果。

结论与展望：我们搭建了高通量计算平台，此平台为高通量计算、材料筛选和预测奠定了坚实的基础。我们研究了12种单原子在38种金属上的吸附以及CO在9056种层状合金上的吸附。利用机器学习方法建立了吸附能预测模型。本工作对后续的高通量计算和催化材料筛选具有重要的意义。

关键词：高通量计算；机器学习；吸附能