Scientists in Singapore have developed software that marks chiral centers on scanning probe microscopy (SPM) images of molecules.
Knowing the chirality of molecules is fundamental for many applications, including heterogeneous catalysis, chiral separation, drug discovery, and anything that relies on surface recognition. The manual determination of the chirality from SPM images can take up to a few days in complex systems, which makes an automated system desirable.
Now, scientists from an artificial intelligence group and an atomic imaging group at the National University of Singapore have worked together to develop a machine learning algorithm that does just that. They tested it on two tightly packed supramolecular aggregates consisting of hexadimethylphenylbenzene and fluorine-substituted hexadimethylphenylbenzene units and found that it accurately identified chiral centers in just a few hours.
SPM image analysis “is currently based on the human ability to ultimately analyze, classify, and interpret subtle variations in STM contrast on the nanoscale,” and is therefore time-consuming and error-prone, explains Jiong Lu, who heads the atomic imaging group . For this reason, the use of machine learning to automatically identify chiral centers would be beneficial.
According to Xiaonan Wang, head of the Artificial Intelligence group, the framework showed “exceptional performance even on non-ideal images with imperfection features … with a limited amount of training data,” suggesting that the model is robust. Several experienced SPM users checked the performance of this model by manually tagging the molecules tested as well. And there is an option to integrate the system with current STM software for convenient use.
SPM is a commonly used technique for analyzing nanostructures. Magalí Lingenfelder, director of the Max Planck EPFL Laboratory for Molecular Nanosciences in Lausanne, Switzerland, has been using SPM for decades and comments, “It’s no surprise that machine learning can have a huge impact on automated image analysis.” Lingenfelder warns, however, that “machine learning for SPM analysis is still in its infancy … I will believe my eyes with every algorithm for a while,” but will gladly accept it as soon as the technology improves.
“If the molecule is prochiral, as in this study, assignments are even more difficult,” says Lingenfelder. The use of algorithms to detect chirality or molecular conformations in complex systems, especially in prochiral systems, could therefore have a major impact. “The revolution is coming,” adds Lingenfelder. “We should welcome it, it should simplify our work.”
Lu and Wang hope this research will result in a model that applies to a variety of molecular systems and is more general. They are now aiming to combine different machine learning systems in SPM instruments to perform other labor intensive tasks that could speed up the material recognition workflow.