2025/04/07 by Tyner, Alexander C. · 1 citation
#Disordered Systems and Neural Networks (cond-mat.dis-nn) #FOS: Physical sciences #Materials Science (cond-mat.mtrl-sci) #Mesoscale and Nanoscale Physics (cond-mat.mes-hall)
paper · doi:10.48550/arxiv.2504.04940
Despite rapid growth in use cases for generative artificial intelligence, its ability to design purpose built crystalline materials remains in a nascent phase. At the moment inverse design is generally accomplished by either constraining the training data set or producing a vast number of samples from a generator network and constraining the output via post-processing. We show that a general adversarial network trained to produce crystal structures from a latent space can be fine tuned through the introduction of advanced graph neural networks as discriminators, including a universal force field, to intrinsically bias the network towards generation of target materials. This is exemplified utilizing two-dimensional topological insulators as a sample target space. While a number of two-dimensional topological insulators have been predicted, the size of the band-gap, a measure of topological protection, remains a concern in most candidate compounds. The resulting generative network is shown to yield novel topological insulators.