Transcript
Solid-state inorganic materials are critical to the growth and development of an electric vehicle, cellphones, laptop batteries, and solar energy technologies. However, finding the ideal materials with the desired functionality for these industries is extremely challenging.
Due to the vast chemical design space and the high sparsity of candidates, experimental trials and first-principle computational simulations cannot be realistically used as screening tools to solve this problem. Instead, researchers have developed a deep learning-based smart algorithm that uses a technique called a generative adversarial network model (or GAN) to improve the material search efficiency by roughly 100 times. Going forward, it has the potential to greatly speed up the discovery of novel functional materials.
The work, published in the journal NPJ Computational Materials, was a collaboration between researchers at the University of South Carolina and Guizhou University.
Inspired by the deep learning technique used in Google’s AlphaGo, which learned implicit rules of the board game Go to defeat the game’s top players, the researchers used their GAN neural network to learn the implicit chemical composition rules of atoms in different elements to assemble chemically valid formulas. By training their deep learning models using the tens of thousands of known inorganic materials deposited in databases such as ICSD and OQMD, they created a generative machine learning model capable of generating millions of new hypothetical inorganic material formulas.
There is almost an infinite number of new materials that could exist, but they haven’t been discovered yet. This algorithm is like a generation engine. Using this model, scientists can generate a lot of new hypothetical materials that have a very high likelihood to exist.
Without explicitly modeling or enforcing chemical constraints such as charge neutrality and electronegativity, the deep learning-based smart algorithm learned to observe such rules when generating millions of hypothetical materials’ formulas.
The predictive power of the algorithm has been verified both by known materials and recent findings in materials discovery literature. One major advantage of this algorithm is the high validity, uniqueness, and novelty, which are the three major evaluation metrics of such generative models.
This is not the first time that an algorithm has been created for materials discovery. Past algorithms were also able to produce millions of potential new materials. However, very few of the materials discovered by these algorithms were synthesizable due to their high free energy and instability. In contrast, nearly 70 percent of the inorganic materials identified by this algorithm appears stable and probably synthesizable.
You can get any number of formula combinations by putting elements’ symbols together. But it doesn’t mean the physics can exist. So, the GAN algorithm and the next step, the structure prediction algorithm, will dramatically increase the speed to screening new functional materials by creating synthesizable compounds.
These new materials will help researchers in fields such as electric vehicles, green energy, solar energy, and cellphone development as they continually search for new materials with optimized functionalities. With the conventional materials discovery process being so slow, the growth of these industries has been limited by the materials available to them.
The next major step for the team is to predict the crystal structure of the generated formulas, which is currently a major challenge. However, the team has already started working on this challenge along with several leading international teams.
Once solved, the two steps can be combined to discover many potential materials for energy conversion, storage, and other applications.
|
