Monolayer films have shown promise as a potential solution for overcoming issues associated with friction and wear in nanoscale devices. Dense layers of surface-bound chains can prevent direct surface-surface contact, thus protecting the surface and reducing frictional forces. The tribological properties of these films have been shown to be closely related to their chemical composition [1, 2], presenting a near-infinite chemical parameter space that can be utilized to optimize tribological properties. In response, complex high-throughput computational screening studies have been enabled through the use of the Molecular Simulation and Design Framework (MoSDeF) [3] and the Signac framework [4]. MoSDeF is a Python-based software suite designed to set up chemical systems in a scriptable manner, while Signac provides a set of tools to create and manage a data workspace. Both software packages allow for a large parameter space to be efficiently explored through non-equilibrium molecular dynamics simulations. Here we examine the tribological effects of terminal group chemistry and film composition, encompassing approximately 36,000 distinct monolayer systems, by far the largest dataset of the tribological properties of monolayer films reported to date. To discover relationships hidden in this large volume of data, machine learning algorithms are employed. Specifically, building upon prior work [1], we use the RDKit library [5], a cheminformatics toolkit, along with the random forest algorithm implemented in scikit-learn [6] to develop a predictive model for the role of terminal group chemistry and film composition on tribological properties. Through this approach we are able to identify which aspects of terminal group chemistry most strongly impact the monolayer tribological properties. For example, it is observed that molecular shape most strongly influences the coefficient of friction, whereby linear and planar molecules are preferable, while charge distribution is most closely linked to adhesion. Compared to previous work, this project explores a much wider parameter space by extending the range of terminal groups examined, adding new backbone chemistries, and examining mixed monolayers. In addition to the random forest algorithm, we also consider other machine learning algorithms and compare their effectiveness in interpreting large volumes of data.
References
[1] Andrew Z. Summers, Justin B. Gilmer, Christopher R. Iacovella, Peter T. Cummings, and Clare McCabe. “MoSDeF, a Python Framework Enabling Large-Scale Computational Screening of Soft Matter: Application to Chemistry-Property Relationships in Lubricating Monolayer Filmsâ€. Journal of Chemical Theory and Computation 2020 16 (3), 1779-1793. DOI: 10.1021/acs.jctc.9b01183
[2] Brewer, N. J.; Beake, B. D.; Leggett, G. J. Friction Force Microscopy of Self-Assembled Monolayers: Influence of Adsorbate Alkyl Chain Length, Terminal Group Chemistry, and Scan Velocity. Langmuir 2001, 17 (6), 1970– 1974, DOI: 10.1021/la001568o
[3] “MoSDeF†[Online]. Available: https://github.com/mosdef-hub.
[4] “signac†[Online]. Available: https://github.com/glotzerlab/signac.
[5] “RDKit Library†[Online]. Available: https://github.com/rdkit/rdkit.
[6] “Scikit-learn†[Online]. Available: https://github.com/scikit-learn/scikit-learn
