Shaking things up: Using machine learning to predict vibrational stability

(Nanowerk News) Countless hours could be saved in the quest to discover new materials thanks to a machine learning program that rapidly and accurately predicts vibrational stability.
Scientists can choose to experiment with millions of theoretical chemical compounds, but many of these proposed materials will never exist in reality, because they are vibrationally unstable.
Information about thermodynamic stability, indicated by formation energy, is readily available in materials databases.
But calculating the vibrational stability of just one untested compound using a conventional approach might require hundreds of hours of processing by a supercomputer, depending on the size of the compound.
Now, researchers in Australia have created a lightning-fast shortcut.
Led by Exciton Science Associate Investigator Dr Sherif Abdulkader Tawfik Abbas of Deakin University, the team took a small subset of approximately 3,100 relatively simple materials from a larger database.
They performed calculations to determine the vibrational stability of these basic materials.
The results were used to train a machine learning model, which is now equipped to accurately predict vibrational stability for a vast array of other proposed materials, including far more complex compounds.
The model produces results with a reasonable degree of accuracy in just seconds.
It could generate important information for researchers working in photonics, medicine superconductivity, energy, and programmable materials, among other applications.
The results, which focused initially on the vibrational stability and structure of inorganic crystals, have been published in Nature’s Computational Materials journal ("Machine learning-based discovery of vibrationally stable materials").
According to Sherif, who is an Alfred Deakin Postdoctoral Research Fellow at the Institute for Frontier Materials, trying to synthesise new materials without knowing their vibrational stability is like building a poorly designed house of cards.
“The formation energy is just half the picture,” he said.
“The other half of the picture is what happens in realistic scenarios when the atoms actually start shaking. We don't know.
“That formation energy doesn't tell us. It only tells us if it is possible hypothetically for the atoms to gather up in space, but it doesn't tell us how they behave together.
“Once they gather up, will the materials stay in that form, or will they collapse?”
Sherif and his collaborators, including Exciton Science Chief Investigator Professor Salvy Russo of RMIT University, worked with Deakin University’s Apply Artificial Intelligence Institute (A2I2) to solve another problem.
A large proportion of the materials in the subset used to train the machine learning model proved to be vibrationally stable, leading to a risk the program would be biased towards stability.
To correct this issue, Sherif and A2I2 artificially inflated the influence of unstable materials during the training process, resulting in a balanced machine learning program able to make impartial predictions.
Source: ARC Centre of Excellence in Exciton Science
podcast
We curated lists with the (what we think) best science and technology podcasts - check them out!