Vast amounts of unstructured structural and functional images are acquired in the quest for scientific discovery. But only a tiny proportion of this data is carefully examined, with an even smaller fraction ever being published. Greater understanding from costly scientific studies which have already been completed is one way to speed up scientific discoveries. Unfortunately, data from scientific tests are usually only available to the person who created them and is familiar with the experiments and guidelines. This issue is aggravated as there were no reliable ways for searching unstructured image datasets for correlations and insight.
An artificial neural network has been recently developed and efficiently trained by a research team at Lehigh University to identify symmetry and structural similarities in materials and create similarity projections. The researchers created an artificial neural network and used machine learning to train it to detect symmetry, patterns, and trends. The researchers utilized this technology for the first time to scan a database of over 25,000 images and correctly classify related elements.
When a neural network is trained, the output is either a vector or a set of numbers that acts as a compact description of the features. These features aid in the classification of objects so that some degree of resemblance can be discovered. But the output may still be huge in space as there might be 512 or more than that different features. Therefore it is compressed into a space that humans can understand, like 2D or 3D.
The researchers extended CNN (Convolution neural network) using a UNET architecture to segment phases5 and nanoparticles6 to extract more information from images. They employed autoencoder structures to unmix statistical data from hyperspectral piezoresponse force microscopy7 and current-voltage spectroscopy8. Furthermore, they used the concept of rotational invariance to extract the orientation of ferroelectric variations from atomically detailed images9.
This neural network has the potential to revolutionize materials research by analyzing massive amounts of data and information from experiments in order to find and interpret patterns in multidimensional data. Scientists and researchers may be able to learn more about the multidimensional structure of materials and the intricacies of structure-property interactions with this artificial neural network. This could reveal previously unnoticed trends and patterns, increase experiment efficiency, and speed up research with better data management and accessibility.