How does machine learning unlock mysteries of quantum physics?

Understanding the complex behavior of electrons has led to findings that transformed society, such as the transistor. Today, with technological advances, electron behavior can be studied far more closely than before, possibly making scientific breakthroughs as changing as the world’s computer. However, the data generated by these tools are too complex to interpret.

A Cornell-led team has developed a way to analyze data from scanning tunneling microscopy (STM) using a machine learning technique that produces subatomic images of electronic motions on material surfaces with varying energies, making data unavailable using other methods.

Some of those pictures took two decades of essential and mysterious materials. You’re wondering what secrets these images contain. We want to unlock those secrets, said Eun-Ah Kim, a physics professor and senior author of Nature June 19 “Machine Learning in Electronic Quantum Matter Imaging Experiments” Research gave new insights into how electrons interact — showing how machine learning can be used to further discover in experimental quantum physics.

At the subatomic level, a sample of billions of electrons interacting with each other and surrounding infrastructure will be present. Electrons’ conduct is partly determined by the tension between their two competing tendencies to move around and stay far away, combined with repulsive energy interaction.

In this study, Kim and the collaborators identified which trends in a superconductive high-temperature material are more important.

Using STM, electrodes are vacuum tuned between the microscope’s leading tip and sample surface and provide detailed information on electrons’ behavior.

The problem is that if you take and record such data, you get image-like data, but it’s not an accurate picture, like an apple or pear. The instrument-generated data is more like a pattern than a traditional measurement curve and 10,000 times more complicated. Kim said we have no excellent tool to study such data sets.

The researchers simulated an ideal environment and additional factors that would lead to changes in electron behavior. She then trained an artificial neural network, a kind of artificial intelligence that can learn the conditions associated with various theories using brain-inspired methods. Inputting experimental data into the neural network, the researchers determined which approaches were most similar to actual data.

This method confirmed the hypothesis that repulsive interaction energy is more effective in electrons’ behavior.

She added that a better understanding of how many electrons interact under different materials and conditions would likely lead to further discoveries, including new material development.

Indeed, the materials that led to the initial transistor revolution were simple materials. Now we can design much more complex elements. If these powerful tools can reveal important aspects leading to the desired property, we wish we could make a material with this property.