Machine Learning Reveals the True Structure of Oxidized Graphene

2020-01-06

What is the actual structure of graphene oxide nanosheets? This question is crucial for optimizing the performance of carbon materials in practical applications, and researchers at CSIRO in Australia are now using machine learning to answer it. They have used over 20,000 possible candidate structures to find a truly representative model, which is significantly different from existing prediction techniques.


Graphene oxide (GO) is a hydrophilic 2D oxidized form of graphene (a single layer of carbon, only one atom thick), with oxygen functional groups that can modify and disrupt the sp2 basal plane of the material, ranging in size from a few nanometers to a few millimeters. The first model of the GO structure was proposed in 1939, suggesting the binding of oxygen to hexagonal carbon sheets with epoxy (1,2-ether) and having the formula C2O. Researchers have been continuously revising this model, considering functional groups that exist in axial binding, which can distort the flat GO structure.In 1998, scientists proposed the Lerf-Klinowski model. In the description of GO, all carbon rings are perfect hexagons, and the out-of-plane deformation caused by functional groups or inherent ripples is basically ignored. Although instructive, this model is quite limited and is largely inconsistent with structures obtained through computational modeling of GO or through electron microscopy images.



Unsupervised Machine Learning Technique


Researchers led by Amanda Barnard at CSIRO Data61 have now re-examined the structure of GO using a new clustering algorithm developed in their laboratory, and predicted the centroid structure that truly represents the material. To extract prototypes, they analyzed the data using an unsupervised algorithm proposed by Cutler and Breiman in 1994.


"In theory, prototype analysis techniques can find points on the convex hull boundary of data clouds in the material feature space," explained the lead author of the study, Benyamin Motevalli. "This means that all possible candidate materials can be described as linear combinations of these prototype (pure) points. The method can even predict prototype structures that are not included in the dataset."



Input Data


Researchers collected input data by creating various sizes and shapes of thin films. They then varied the oxygen concentration in the films and added different chemical groups, which were distributed in different ways.


These samples included hydroxyl, ether, double bond, alkyl (cyclohexane) groups, as well as out-of-plane distortions (caused by defects) that are clearly beyond the Lerf-Klinowski model.


The team included four different film morphologies: hexagonal (49.5%), triangular (14.3%), rectangular (30.5%), and rhombic (5.7%). The total number of atoms in each sample ranged from 191 to 1949, including C, H, and O atoms.


Motevalli explained, "The density and distribution of oxygen functional groups play a crucial role in obtaining GO properties. Therefore, for 24 original graphene oxide nanosheets, we sampled many O/H concentrations, each with hundreds of random distributions." The O/C ratio ranged from 4.05% to 52.08%, and the H/C ratio ranged from 2.22% to 49.26% in each case.



28 Structures Substitute 20,396 Samples


Using this approach, researchers identified three representative GO nanosheets that are "average" structures in a 223-dimensional space.


They also said that they identified 25 "pure" GO nanosheet structures that capture all the complexity and diversity of the entire dataset of 20,396 samples. These 25 structures can be used as linear combinations to represent the entire set.


Motevalli said, "These 28 structures (25 structures and 3 genotypes) together can substitute for 20,396 samples without losing any information. They can also be used as single model structures with the correct chemical composition."



Eliminating Guesswork and Bias


He explained, "Our collection of 20,396 GO nanosheet structures required years of work and over 30 million core supercomputer hours to generate electronic structures. Simplifying this collection to 28 most important structures will enable other research groups to make representative and reliable GO predictions within this timeframe."


He added that this method also eliminates guesswork and bias in GO computational models and provides the consistency required for benchmarking. "If all researchers involved in GO research use the same model structures, we can easily compare and correlate results from laboratories around the world."


The researchers plan to use supervised machine learning to explore the relationship between GO structures and properties, and predict how different types of samples will perform under different conditions and applications. "Examples include charge transfer characteristics, or studying the role of defects and distortions and how they affect fault tolerance," said Motevalli.

RECENT AI NEWS

RECENT AI TOOLS