Machine Learning Cascade for Defect Detection in Graphene

Sammanfattning

Graphene is a two-dimensional material with exceptional mechanical, electrical, and thermal properties, but its performance is highly sensitive to structural defects that arise during production. This thesis investigated whether a machine learning model could determine the type, orientation, and position of structural defects in graphene from a small number of topological probe measurements, without directly observing the defect. Graphene was modelled as a graph, where carbon atoms are represented as nodes and bonds as edges, and four defect types were considered: single vacancy, double vacancy, Stone-Wales, and 555-777. The measurement signal exploited the fact that the perfect graphene lattice contains no odd-length cycles, while defects introduce odd-length cycles locally whose length grows with distance from the defect. A synthetic dataset of 100,000 samples was generated using an adaptive probing strategy that placed six probes per sample, guided by breadth-first search to estimate the defect location. A three-stage cascade of machine learning models was then trained to predict defect type, orientation, and position sequentially. The cascade achieved a defect type accuracy of 96.4%, an average orientation accuracy of 88.2%, and a mean localisation distance of 0.283 lattice nodes, outperforming a flat baseline classifier by a factor of 4.5 in localisation accuracy. The results show that topological measurements alone carry sufficient information for accurate defect classification and localisation in graphene.