Determining the structure of functionalized graphene for tailored thermomechanical properties using ML techniques

Chemical functionalization of graphene with various chemical groups unlocks an infinite number of variations for nanosheet design modifications. However, the prohibitive cost of molecular dynamics simulations and the overwhelmingly large number of design variables render the inverse design problem intractable when conventional approaches are used. To this end, we develop an MD-powered, data-driven framework to enable fast and accurate identification of the layout that exhibits a given set of user-prescribed thermomechanical properties. Specifically, we generate a dataset with 1200 records, combining the layout and thermomechanical properties (Young's modulus, thermal conductivity, maximum stress and strain at maximum stress) of functionalized graphene sheets with hydrogen and methyl groups of appropriate coverages. A variety of regression models using Label and Bag-of-Words encoding were trained with Support Vector Regression, Ridge Regression and Gaussian Process Regression models showing best predictive performance, with considerably high values for the corresponding coefficients of determination (R ² > 0.9 for thermal conductivity, Young's modulus and maximum stress) on a hold-out test set, with mean absolute percentage error (MAPE) remaining below 1% in most cases. Finally, an evolutionary optimization process, in tandem with the trained Machine Learning (ML) models, was employed for finding graphene layouts that possess a set of user-defined target properties. MD-validations of the obtained designs confirmed the applicability of the approach while revealing acceptable deviations for thermal conductivity values and even better alignment for the mechanical properties. In summary, the proposed approach succeeds in a 7 orders of magnitude speedup in estimating the thermomechanical properties of functionalized graphene sheets when compared to pure MD simulations, and up to 6 orders of magnitude faster identification of layouts with prescribed properties, benchmarked on a nanosheet (220 × 100 Å) with 8528 atoms using a 64 core AMD EPYC workstation.