AI has learned to generate the molecular structure of materials
We are on our way to creating an AI microscope that will find interesting objects without knowing what to look for.
Artificial intelligence capable of mimicking human creativity can now generate artificial scientific data, a new step towards fully automated data analysis.
Scientists at the University of Illinois at Urbana-Champaign created artificial intelligence which generates synthetic data based on the results of microscopic experiments commonly used to study the structure of materials at the atomic level.
Using technologies similar to those used in generative AI, the created neural network is able to add background noise and experimental imperfections to the generated data, which speeds up and simplifies the identification of material properties.
Images processed by CycleGAN retain the types and positions of atomic defects from the original simulated images
In materials science, various methods of artificial intelligence and machine learning are already actively used to analyze data, but these methods require constant and costly human participation. Increasing the efficiency of these procedures involves the use of a large amount of labeled data, which allows the program to better understand what it needs to look for. However, the data must account for a wide range of background noise and experimental inaccuracies, making it difficult to model.
Since collecting and labeling such a volume of data using a real microscope is an impossible task, scientists have developed a generative AI that can create a large amount of artificial training data from a relatively small set of real labeled data.
Two microscopic images of the surface of the material. The left image was created by AI and the right image was taken with a microscope
To do this, the researchers used a cyclic generative adversarial network, or CycleGAN, which can transfer image style from one domain to another without using paired examples.
The general idea of adversarial networks is based on a “competition” between two networks: a generator and a discriminator. The generator tries to create artificial data that looks like real data, and the discriminator tries to distinguish this artificial data from the real one. Over time, both networks get better at their tasks, leading to the creation of high-quality artificial data that is almost indistinguishable from the real thing.
By training on a small sample of real microscopic images, the AI was able to generate images that were used to train the analytical algorithm. Scientists can now recognize a wide range of structural features in a material despite background noise and systematic imperfections.
The importance of the study is that the experts did not have to train the AI in advance to recognize background noise or image distortions in the microscope. This means that even if there is something that scientists did not think about, CycleGAN is able to learn it on its own and use it in its work.
The research team has incorporated CycleGAN into their experiments to detect defects in 2D semiconductors, a class of materials that are easy to apply in electronics and optics but difficult to characterize without the help of AI. However, this method has much broader application possibilities.
Scientists noted that with the help of technology it is possible to create a self-controlled (autonomous) microscope. However, to create such a microscope, the biggest obstacle was understanding how to process the data, but the work done by the experts solves this problem. According to the experts, they were able to show how you can teach a microscope to detect interesting things without having an idea of what exactly to look for.