An assistant professor of materials science and engineering at Johns Hopkins has developed an artificial intelligence framework that can map the electrons for 65 elements on the periodic table and their combinations, accelerating the development of better computer chips, solar cells, LED lights, and more. How his creation, called SlaKoNet, works is explained in The Journal of Physical Chemistry Letters.
“We needed faster predictions for electronic movement in materials, which will determine how they move in small chips and ultimately improve how chips are made,” says Kamal Choudhary, who holds a joint appointment in the Department of Electrical and Computer Engineering. “Technology chips include many layers of materials, and we need to know how the electrons will flow and if adding an element to the chip would make them flow faster. This model can accurately predict that for those 65 elements.”
Instead of creating something new, Choudhary used AI to upgrade the long-standing Slater-Koster tight-binding formalism, a method for predicting electronic band structures. These structures control electrical behavior by allowing energy from electrons to flow within a material.
“The Slater-Koster tight-binding formalism requires manually set parameters to predict how electricity will flow through a material, which is tedious and prone to error,” says Choudhary. “SlaKoNet is a framework that modernizes the classical method, modeling combinations of electrons to discover which blend performs the best.”
The model retains its basis in physics but uses a neural network framework—a type of artificial intelligence that recognizes patterns and makes decisions to optimize parameters of prediction equations. SlaKoNet learns how electrons behave in many metals, semiconductors, and insulators, easily adapts to new datasets, and is eight times faster than the standard central processing unit (CPU) calculation method.
“It was trained on large datasets, particularly the JARVIS-DFT TBmBJ dataset from the National Institute of Standards and Technology, because it provides a highly accurate foundation for learning,” he says. “In total, the model was trained on roughly 20,000 material combinations across 65 elements, including oxides, carbides, nitrides, halides, and intermetallics. This breadth ensures that its predictions remain reliable across a wide range of materials.”
SlaKoNet currently includes most materials that are used in quantum and semiconductor technologies, but Choudhary plans to extend it to the full periodic table. “The long-term goal is to make a universal model for all elements and their combinations,” says Choudhary.
Now, the model is live on Choudhary’s ChatGPT platform for materials scientists called atomgpt.org. “My ultimate goal is to make everything open access for research, so SlaKoNet is freely available for scientists in academic and industry environments,” he says.
Choudhary used computational materials available at Johns Hopkins University to complete this project.