(415c) Genetic Algorithm Enhanced By Atomistic Neural Network: Pt Clusters at the H2 Atmosphere as an Example | AIChE

(415c) Genetic Algorithm Enhanced By Atomistic Neural Network: Pt Clusters at the H2 Atmosphere as an Example


Sautet, P., University of California Los Angeles
Revealing the catalyst structures at the atomic level is essential to rationally improve or design new efficient catalyst. Accordingly, it is a key goal for theoretical catalysis studies. Although theoretical calculations can reveal the catalyst structures in atomic details, which is beyond the capability of most current experimental techniques, there are still some challenges. First, theoretical calculations based on first-principles calculations are very expensive, therefore the simulations are normally limited within several hundred atoms; Secondly, realistic catalyst structures are quite diverse and they are easily affected by reactants, products et al1. The catalyst might be associated to several competing low energy structures, and not to a single unique structure, which in turn needs more computations. Finally the exploration of potential structures on a system with a large number of degrees of freedom is computationally complex2. All of those factors hamper the understanding of the relationship between reaction activity and catalyst structures.

We proposed two strategies to address above obstacles. Firstly, in order to investigate all the possible catalyst structures, we devise a scheme to enhance the capability of normal genetic algorithm (GA)3. Considering that the most stable structure is not necessary the most active catalyst structure, a comprehensive exploration of many low energy catalyst isomers is vital. With this in mind, our enhanced GA is aimed at finding all the possible structures that could have a reasonable population in catalytic temperatures rather than only locating the global minimum. Secondly, since the genetic algorithm will exploit a large number of structure optimizations, it is time-consuming to carry out all optimizations by DFT directly. Therefore, we accelerated the computationally expensive structure optimizations with an atomistic neural network (ANN)4, which is trained before performing GA optimizations. An iterative scheme is also utilized to efficiently find the optimal training set structures during the ANN training procedures.

With those tools, we studied the structures of platinum clusters covered with hydrogen adsorbates (PtnHx). For every composition, we can find all the possible structures whose energy lies between the Egm (gm=global minimum) and Egm+0.5 eV. Those structures cover all the catalytically relevant structures that may appear during catalytic reactions, therefore we can locate the most important catalytic species from the structure population. The influence of the supports is in consideration. In the end, we can improve our understanding of Pt clusters at more realistic conditions.

All in all, we presented a combined routine, which takes advantages of two cutting-edge techniques: global optimizations and neural-network methods. This method can efficiently find out the low energy structures for a large system without loss of energy accuracy vs DFT. Then the method is utilized for studying a realistic system-Pt cluster at the atmosphere of H2. Besides the example in this study, this scheme can be easily applied to similar catalytic problems benefiting from the generality of genetic algorithm and atomistic neural network.

1. Mager-Maury, C.; Bonnard, G.; Chizallet, C.; Sautet, P.; Raybaud, P., H2-induced reconstruction of supported Pt clusters: metal-support interaction versus surface hydride. ChemCatChem 2011, 3 (1), 200-207.

2. Zhai, H.; Alexandrova, A. N., Ensemble-Average Representation of Pt Clusters in Conditions of Catalysis Accessed through GPU Accelerated Deep Neural Network Fitting Global Optimization. J Chem Theory Comput 2016, 12 (12), 6213-6226.

3. Vilhelmsen, L. B.; Hammer, B., A genetic algorithm for first principles global structure optimization of supported nano structures. J Chem Phys 2014, 141 (4), 044711.

4. Artrith, N.; Hiller, B.; Behler, J., Neural network potentials for metals and oxides - First applications to copper clusters at zinc oxide. Physica Status Solidi (b) 2013, 250 (6), 1191-1203.