Machine Learning Suggests Possible Bias in Molecular Simulations for Asphaltene Systems

Abstract

Asphaltenes are defined as a solubility class rather than a singular family of molecular compounds; therefore, there is no unique molecular structure that corresponds to asphaltene. Perhaps as a consequence of this definition, direct probing of the physical and chemical properties of asphaltene molecules has been historically challenging in the laboratory. It was only in the past few years that experimental imaging of asphaltene structures resolved representative molecular geometries. Could the identification deriving from these experiments be useful for moving forward our collective understanding of the properties of asphaltenes? Is the experimental identification of numerous asphaltene molecules implying that simulations should consider all of these compounds for providing useful insights? Is it possible that the community has been considering a limited sample of asphaltene molecules when attempting to shed light on asphaltene behavior? This review attempts to address these questions with the help of unsupervised clustering. The results demonstrate that the experimental asphaltene structures can be clustered into four groups that resolve different physical properties of asphaltenes. Then, we probe the existing literature to assess whether compounds representative of the molecules experimentally identified as "asphaltenes" using noncontact atomic force microscopy (nc-AFM) [