Die DPG-Frühjahrstagung in Dresden musste abgesagt werden! Lesen Sie mehr ...
Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe
O: Fachverband Oberflächenphysik
O 42: Focus Session: Innovation in Machine learning PRocEsses for Surface Science (IMPRESS)
O 42.7: Vortrag
Dienstag, 17. März 2020, 12:30–12:45, TRE Phy
Chemicaly reasonable models for automatic interpretation of AFM images — •Prokop Hapala1, Niko Oinonen2, Fedor Urtev2, Benjamin Alldritt2, Ondrej Krejci2, Filippo F. Canova2, Fabian Schulz2, Juho Kannala2, Peter Liljeroth2, and Adam S. Foster2 — 1Dep. Condensed Matter Theory, FZÚ AV ČR — 2Dep. Applied Physics, Aalto University
During the last year we pioneered machine-learning methods for reconstruction of molecular structure from high-resolution AFM images of non-planar organic molecules [1], which opens the way to broader application of this experimental technique for single-molecule analysis [2] e.g. in the pharmaceutical industry. Nevertheless, a scheme relying purely on general-purpose image recognition tools (such as convolutional neural networks) is sub-optimal as it discards physical insight. Incorporation of physical models and chemical intuition (e.g. bonding topology of carbon) into the scheme could considerably regularize the model, thus making it more reliable in situations when input information is limited. This is especially important, since the AFM provides rather limited information about deeper molecular structure which does not come into direct contact with the tip. The challenge is to formulate a model which encode relevant rules of organic chemistry, including both atomic and electronic structure (such as *electron force-field* [3]), while it is simple enough to be conveniently embedded into machine learning framework,[1]B.Alldritt,et.al.,Science Advances,(2019)accepted,[2]B.Schuler,et.al.,JACS,137(31),9870-9876,(2015),[3]H.Xiao,et.al.,Mechanics of Materials,90,243-252,(2015)