Deep-learning algorithm generates viable hypothetical materials
Image credit: Dreamstime
University of South Carolina researchers have used deep-learning techniques to create a model capable of generating millions of hypothetical inorganic material formulae.
Solid-state inorganic materials are vital for the growth of many emerging technologies and their development has historically been fuelled by the demands of industry. Today, these materials are used in electric vehicles, mobile device batteries and solar panels.
However, finding materials with the perfect characteristics for these purposes is a challenge. Due to the vast possibilities in the chemical design space and the extreme sparsity of candidates, experimental trials and traditional computer simulations cannot practically be used as screening tools.
Now, computer scientists from the University of South Carolina and their collaborators have developed a deep learning-based algorithm which uses a generative adversarial network model to improve material search efficiency by up to two orders of magnitude.
The researchers used the network to extract the implicit chemical composition rules of atoms in various elements and assemble chemically valid formulae. This means that – without explicitly modelling chemical constraints – the algorithm learned to observe the rules of chemistry.
Then, by training the model using the tens of thousands of known inorganic materials, they created a generative model capable of generating millions of new and viable (charge neutral and electronegativity-balanced) inorganic material formulae.
“There is almost an infinite number of new materials that could exist but they haven’t been discovered yet,” said Professor Jianjun Hu. “Our algorithm, it’s like a generation engine. Using this model, we can generate a lot of new hypothetical materials that have very high likelihoods to exist.”
Guizho University’s Professor Shaobo Li, who was also involved in the study, commented: “Our major advantage of our algorithm is the high validity, uniqueness and novelty, which are three major evaluation metrics of such generative models.”
This is not the first algorithm to have been created for the purposes of materials discovery, although very few of the suggested materials generated by previous algorithms were viable due to complications with high free energy and instability. In contrast, nearly 70 per cent of the materials generated by this new algorithm were judged stable and possibly synthesisable.
“You can get any number of formula combinations by putting elements’ symbols together but it doesn’t mean the physics can exist,” said the University of Southern Carolina’s Professor Ming Hu. “So, our algorithm and the next step, structure prediction algorithm, will dramatically increase the speed to screening new function materials by creating synthesisable compounds.”
The researchers hope that their work could contribute to speeding up the timeline of materials discovery, boosting the growth of industries such as solar energy and electric vehicles.
Sign up to the E&T News e-mail to get great stories like this delivered to your inbox every day.