Physics Guided Deep Learning for Generative Design of Crystal Materials with Symmetry Constraints

Abstract

Discovering new materials is a long-standing challenging task that is critical to the progress of human society. Conventional approaches such as trial-and-error experiments and computational simulations are labor-intensive or costly with their success heavily depending on experts' heuristics. Recently deep generative models have been successfully proposed for materials generation by learning implicit knowledge from known materials datasets, with performance however limited by their confinement to a special material family or failing to incorporate physical rules into the model training process. Here we propose a Physics Guided Crystal Generative Model (PGCGM) for new materials generation, which captures and exploits the pairwise atomic distance constraints among neighbor atoms and symmetric geometric constraints. By augmenting the base atom sites of materials, our model can generates new materials of 20 space groups. With atom clustering and merging on generated crystal structures, our method increases the generator's validity by 8 times compared to one of the baselines and by 143% compared to the previous CubicGAN along with its superiority in properties distribution and diversity. We further validated our generated candidates by Density Functional Theory (DFT) calculation, which successfully optimized/relaxed 1869 materials out of 2000, of which 39.6% are with negative formation energy, indicating their stability.

Data-driven deep generative design of stable spintronic materials

“…26,29 Instead, VAE models always generate crystal structures with very low symmetry. 28 On the other hand, we only use materials falling in the three space groups Fm3 ¯m, F4 ¯3m, and Pm3 ¯m because these three space groups are with the greatest number of materials in the OQMD 30 using selection criteria in CubicGAN.…”

“…The generator takes random noise as input to generate fake samples and the discriminator tells the fake samples from the real ones. CubicGAN 24 and PGCGM 28 are two typical crystal generative models using GANs. Both are provided with the space group in the training and physical losses are added in PGCGM to improve the performance.…”

Data-Driven Deep Generative Design of Stable Spintronic Materials

“…The generator takes random noise as input to generate fake samples and the discriminator tells fake samples from real ones. CubicGAN [17] and PGCGM [21] are two typical crystal generative models using GANs. Both are provided with the space group in the training and physical losses are added in PGCGM to improve the performance.…”

“…3) We are not trying to generate crystal structures in 230 space groups like what VAE models claim [19,22]. Instead, VAE models always generate crystal structures with very low symmetry [21]. On the other hand, we only use materials falling in three space groups of Fm 3m, F 43m, and Pm 3m because these three space groups are with greatest number of materials in OQMD [23] using selection criteria in CubicGAN.…”

Generative adversarial networks and diffusion models in material discovery

“…Discovering new structures in an efficient and effective way requires a method of processing enormous amounts of data, quickly identifying patterns that are present within the dataset, and extrapolating those patterns outside of the dataset so that new materials can be discovered. With these constraints, deep learning-based generative modeling has shown an immense amount of promise within the field of materials discovery [4,5,6,7,8,9,10,11].…”