Exploring self-assembly using machine learning, STM, and toy models
COMP Seminar (Otakaari 1). Speaker: Dr. Filippo Federici Canova.
Map © OpenStreetMap. Some rights reserved.
The talk summarises all projects — big and small — undertaken while working at AScI/COMP since 2014.
By far the simplest, and perhaps most rewarding, was the study of self-assembled BTB (1,3,5-benzenetribenzoic acid) [1]. Scanning probe measurements show the formation of a regular honeycomb lattice, and peculiar deformation after depositing phthalocyanine. Using a simple toy model, we were able to understand the origin of the deformation pattern.
A fierce debate over the structure of self-assembled graphene nanoribbons (GNRs) [2] was settled after scanning probe measurements at Aalto. Moreover, these studies have given a lot more insight into the reaction process, which can be used to intelligently design GNR precursors [3].
Machine learning strategies have been developed and applied to other projects, including the main one [WHICH ONE?]. Under its cool name, machine learning obscures a set of statistical tools for finding patterns and fitting complex data. Like all tools, sometimes they are not right for the task, or they break. The talk shows a few examples of machine learning applied to fitting classical force fields from ab-initio data [4], and explains the most common reason preventing their success (for now).
Finally, the talk presents the latest results of machine learning methods applied to the prediction of lubricant performance [5] and our attempts to extend the framework to quantum chemistry.
[1] J. Phys. Chem. C, 120, 8772 (2016)
[2] ACS Nano, 8, 9181 (2014)
[3] J. Phys. Chem. C, Just Accepted
[4] J. Comp. Chem., 36, 1187 (2016)
[5] J. Chem. Theory Comput., 13, 3 (20