Structure prediction of two-dimensional materials based on neural network-driven evolutionary technique
Introduction
These days, making an attempt to find a perfect combination of chemicals suitable for specific applications, a chemist (or a materials engineer) has two options. The first one is to conduct an experiment, synthesize a sample, and then characterize it with proper methods. The second option is to use a theoretical study first, which may reveal properties of a given material with very high credibility. This is possible mainly due to the huge success of the ab-initio methods, based on Density Functional Theory (DFT) [1], which is currently one of the main computational tools for materials sciences, used at the atomic level. Increasing computational power and data storage capacity have made possible the calculation and analysis of properties for a huge amount of materials. That high throughput (HT) studies (e.g. [2]) are becoming more and more popular because at relatively low cost of these allow scientists to pick out structures with desirable properties from a large group of average ones. Moreover, large amounts of generated data inspired studies based on methods used earlier in pure data science, such as deep learning methods [3], usually based on artificial neural networks (NN) [4] and evolutionary optimization methods [5]. An additional effect of these HT studies is the creation of free-accessible on-line databases [6], [7], containing thousands of calculated structures.
From the implementation point of view of the previously mentioned methods, one may observe three main directions. First one is the application of the methods based on evolutionary algorithms, as implemented for example in XtalOpt [8] or USPEX codes [9]. Especially the latter was extremely successful, being used in more than half a thousand studies. The second direction one may take is a swarm optimization [10], implemented for example in the CALYPSO code [11]. A third way is the use of (more and more popular in all fields of science) artificial neural networks.
A machine learning (ML) approach is today a well established branch of materials science [12]. These studies take advantage of the large amount of available data to train the network and then use it to predict properties of yet unknown structures. For example, a thorough studies were conduced for the structure prediction of binary [13] and ternary compounds [14]. Also, in the study [15], a ML model has been trained to discover Heusler compounds with a true positive rate of 0.94. Moreover, according to [16], it is possible to build a NN model, which predicts crystals’ properties with greater accuracy (with respect to an experiment) than a Hybrid DFT method does. One of the disadvantages of this approach, though, is the necessity of having the proper set of data, on which the NN would learn.
It is also possible to combine methods belonging to one of the aforementioned paths. Especially promising (e.g. in terms of performance) seem to be methods combining NN and evolutionary approach [17]. In these hybrid methods, NN and DFT evaluations of the total energy are used in the pursuit of the optimal structure.
Low-dimensional materials play an increasingly important role in the pursuit of the new generation of structures that will build future logic systems. Extensive theoretical work on this subject work resulted in significant number of potentially valuable substances. Similarly, significant progress has been made in the experimental field [18]. Also, very recently, the first two-dimensional (2D) magnetic material has been experimentally confirmed [19]. These materials are not only flat as graphene but may also be built of several monoatomic layers [18]. Also, experimentally, a 2D layer is always placed on a substrate, so theoretical models have to include its impact.
In this paper, we present a simple, yet effective algorithm of finding new two dimensional structures, either in vacuum or on a substrate, starting only with its stoichiometry. Since the main issue of the new materials search problem is a vast search space [9], to effectively reduce this space, the algorithm combines trained NN and an evolutionary approach.
Section snippets
Description of the used algorithm
This section presents the detailed algorithm of an approach, that we developed to predict properties of 2D structures.
Implementation
The above algorithm has been implemented in the NGOpt code (Neural network Genetic algorithm Optimizer), written entirely in the Python programming language [24]. The code uses a few external free libraries.1
The main data structure is the Individual class, which inherits from the Atoms class of the Atomic Simulation Environment (ASE) [25] and describes a distinct structure. The initial population (a set of Individuals) is
Tests and examples
As examples three systems have been chosen, each having a little different nature. First, though, the accuracy of the NN model should be examined.
Conclusions
We have developed and implemented an algorithm of 2D structure prediction. The method, based on a combination of artificial neural networks and evolutionary approach has been tested for the cases of simple monolayer materials (MoS2), layered materials (CoGa2S4) as well as 2D multilayers on a substrate (boron trilayer on SiC). In all the cases, the final structure was consistent with previous literature results. Finally, the code has been tested for the case of unknown structure, which resulted
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Acknowledgments
Numerical calculations were supported in part by PL-Grid Infrastructure.
References (46)
- ...
Acta Chem. Scand.
(1970)- et al.
High Throughput Computational Screening for 2D Ferromagnetic Materials: The Critical Role of Anisotropy and Local Correlations
2D Mater.
(2019) - et al.
Inhomogeneous Electron Gas
Phys. Rev.
(1964) - et al.
The Computational 2D Materials Database: High-Throughput Modeling and Discovery of Atomically Thin Crystals
2D materials
(2018) - Deep Learning | The MIT Press https://mitpress.mit.edu/books/deep-learning (accessed Aug 19,...
- Neural Networks in Materials Science...
- Handbook of Evolutionary Computation, Lslf edition.; Back, T., Fogel, D. B., Michalewicz, Z., Eds.; Oxford University...
- ...
- ...
Crystal Structure Prediction Using Ab Initio Evolutionary Techniques: Principles and Applications
J. Chem. Phys.
Swarm Intelligence
Classifying Crystal Structures of Binary Compounds AB through Cluster Resolution Feature Selection and Support Vector Machine Analysis
Chem. Mater.
Disentangling Structural Confusion through Machine Learning: Structure Prediction and Polymorphism of Equiatomic Ternary Phases ABC
J. Am. Chem. Soc.
High-Throughput Machine-Learning-Driven Synthesis of Full-Heusler Compounds
Chem. Mater.
Prediction Errors of Molecular Machine Learning Models Lower than Hybrid DFT Error
J. Chem. Theory Comput.
Genetic Algorithms for Computational Materials Discovery Accelerated by Machine Learning
npj Comput. Mater.
2D Materials: The Thick and the Thin
Nat. Rev. Mater.
Layer-Dependent Ferromagnetism in a van Der Waals Crystal down to the Monolayer Limit
Nature
Cited by (2)
When Machine Learning Meets 2D Materials: A Review
2024, Advanced Science