Two-Stage Machine Learning Framework for Accurate Discrimination of Isomers and Very-Similar Molecules on Surfaces

Area of Science:

• Surface Science
• Nanotechnology
• Computational Chemistry

Background:

• Accurate detection of on-surface organic isomers and similar molecules is vital for understanding chemical reactions.
• Current methods lack efficiency, robustness, and automation for distinguishing closely related molecules on surfaces.

Purpose of the Study:

• To develop an automated, precise framework for detecting and identifying multiple types of sparsely distributed molecules on surfaces.
• To address the challenge of distinguishing molecules with analogous features in scanning tunneling microscopy (STM) images.

Main Methods:

• Implemented a two-stage convolutional neural network (CNN) framework named ReSTOLO.
• Utilized YOLO v5.m for molecular localization and ResNet-101 for classification, connected via box normalization.
• Trained the models on an augmented STM image database of single molecules.

Main Results:

• Achieved approximately 20% improvement in precision, recall, and accuracy compared to conventional methods.
• Demonstrated effectiveness in analyzing a surface reaction with six molecules exhibiting nearly identical STM signatures.
• Showcased robust capabilities in automatically pinpointing and discriminating between molecular species with similar configurations.

Conclusions:

• ReSTOLO offers an efficient, automated solution for precise molecular discrimination in complex surface reactions.
• This framework significantly advances the field of STM tip-manipulated surface chemistry.
• The approach facilitates detailed analysis of surface reaction mechanisms and molecular properties.