A group of scientists used machine learning (ML) to precisely identify the atomic structures of interfacial ionic hydrates based on AFM images, including the orientation of water molecules as well as the position of each atom. Moreover, they observed that using a neural network trained with interfacial ionic data, the structure prediction of water hydrates in a cost-effective way is possible. Therefore, this particular study provides an economical and efficient methodology to determine the atomic structures of complex systems from AFM images as well as help to interpret other sophisticated experimental results.
Background and Limitations of Current Method
The atomic structure information of ionic hydrates is critical for understanding the unique physical and chemical features of the ionic hydrate/solid interface. Nowadays, to characterize interfacial water structure at an atomic level, qPlus-based AFM is used. However, the real problem with the current technology is that when using AFM in sharp contrast to oxygen atoms, the hydrogen atoms are not visible. Therefore, applying conventional methods generally requires trial and error processes to omit a huge number of less probable structural models with different OH orientations.
Summary
A group of scientists developed an ML method with transfer learning, which could economically determine the structure of interfacial ionic hydrates from AFM images. For this purpose, they took Na+ hydrates as an example. Firstly, they trained a neural network structure (NN) with many simulated AFM images of interfacial water structures through classical MD. After that, they retrained the NN based on DFT-computed electrostatic potentials. By doing so, the NN was able to achieve a prediction accuracy of 95% for both oxygen and sodium as well as 85% for hydrogen.
More from AZoM: Trends in Atomic Force Microscopy
The orientation of water molecules and the position of each atom can be easily identified from the NN’s predicted representation of the designed structure. The efficiency and accuracy of this prediction that was achieved by using this method are not only far superior to that of the trial and error process but this high efficiency and accuracy have not been achieved by any prior machine learning methods. Moreover, the cost-effective machine learning method is also a general workflow for predicting structures using AFM images, and it is possible to extend it to other complex systems like the surface of ice, as well as to deduct information from findings of other spectroscopic experiments.
Findings
The whole process goes through various phases including data preparation for preliminary training, interfacial water structure prediction on the basis of preliminary training, Na+ hydrates’ structure prediction based on transfer learning, and then validity assessment of the transfer learning.
For this final phase, different training sets with different data volumes are used to train NN with or without transfer learning. The prediction accuracy of each set is then calculated by a dataset containing 500 hydrate structures. NN trained with transfer learning shows a better prediction accuracy of all atomic species as opposed to transfer learning for N more than 500. The hydrogen prediction accuracy based on a natively trained NN without transfer learning is nearly zero.
However, higher prediction accuracy is achievable with just a few thousand data using transfer learning of primarily trained NN. Also, an important observation is that when N is smaller than 500, the directly trained NN perform better than transfer learning in Na+ predictions. This can be because of two factors. One is due to the stronger signal for Na+ in AFM images that makes learning easy for NN, and the second is that to retain primarily trained NN through transferred learning, enough Na+ hydrate data is required to eliminate previous local minima.
Moreover, this apparent limitation of transfer learning soon vanishes when N is increased to 500. The prediction accuracy of transfer learning, however, shows no improvement when N is higher than 1000, which indicates that in this case, there are only a few hundred data necessary to train a precise and accurate NN.
We can conclude that NN trained primarily with interfacial water data is transferable through transfer learning to ionic hydrate systems, which enables the training of NN with very limited data to obtain very high precision predictions even for complex structures.
Future Implications
This study is of great importance for AFM image-based high-resolution atomic structure predictions. However, further improvements can be made. For example, experimental errors like tip drifting and noise distribution need more careful handling in order to improve prediction accuracy.
In the future, by introducing NN architectures, this method will be able to determine 3D structures like ice. This machine learning method with transfer learning can be further modified for a variety of applications; for instance, interpreting other experimental measurements like TEM, SEM, and STM.
References and Further Reading
Tang, B., Song, Y., Qin, M., Tian, Y., Cao, D., Wu, Z. W., ... & Xu, L. (2022). Machine learning aided atomic structure identification of interfacial ionic hydrates from atomic force microscopy images. arXiv preprint arXiv:2203.12443. https://arxiv.org/abs/2203.12443
Disclaimer: The views expressed here are those of the author expressed in their private capacity and do not necessarily represent the views of AZoM.com Limited T/A AZoNetwork the owner and operator of this website. This disclaimer forms part of the Terms and conditions of use of this website.