Desenvolvimento de modelo computacional baseado em machine learning para predição de propriedades anticorrosivas de compósitos poliméricos aditivados com grafeno

Abstract

The direct and indirect damages caused by the corrosive phenomenon in the deterioration of metals represent costs and consequences that are often incalculable, affecting not only the public, commercial and industrial sectors, but also posing environmental and public health problems. An approach that has increasingly attracted the interest of the scientific and industrial community is the use of polymeric nanocomposites with anticorrosive properties that act as a barrier between the metallic substrate and the corrosive environment. These nanocomposites are developed from the combination of a polymer matrix and nanoparticles (dispersed phase) with desirable properties, resulting in a material with optimized properties. However, due to the range of possible combinations between compounds in different concentrations, it becomes impractical to analyze by conventional methods which combinations are ideal for a specific application. To circumvent this problem, one can resort to the use of predictive models based on machine learning. With this in mind, this work investigated computational models based on machine learning for predicting the anticorrosive properties of polymeric nanocomposites. 100 literature articles were cataloged, resulting in a data set of 40 instances. Regression models to predict the corrosion potential of covered carbon steel with selected nanocomposites were trained through the Automated Machine Learning (AutoML) approach, using chemoinformatics tools to adjust the representation of the constituent polymers of each composite. When the response variable for the test model was represented by the corrosion potential, a relatively low evaluation metric value was obtained for the mean quadratic error (𝑅 2 = 0,143). However, a better result could be observed for the mean quadratic error (𝑅 2 = 0,429) in the test model when the response variable is represented by the difference between corrosion potentials, with and without the presence of graphene oxide in the nanocomposites.