Hundreds of new, stable, one-dimensional materials from a generative machine learning model

Abstract

We use a generative neural network model to create thousands of new one-dimensional (1D) materials. The model is trained using 508 stable one-dimensional materials from the Computational 1D Materials Database (C1DB) database. More than 500 of the new materials are shown with density-functional theory calculations to be dynamically stable and with heats of formation within 0.2 eV of the convex hull of known materials. Some of the new materials could also have been obtained by chemical element substitution in the training materials, but completely new classes of materials are also produced. The band structures, electronic densities of states, work functions, effective masses, and phonon spectra of the new materials are calculated, and the data are added to the C1DB.

Inverse Design of Next-Generation Superconductors Using Data-Driven Deep Generative Models

“…To circumvent this obstacle, generative machine learning algorithms can be used to design new candidate structures. With the popularity of generative tools such as DALL-E, which uses deep learning models to generate digital images from natural language descriptions (prompts), the interest in using deep generative models for scientific applications has immensely increased. − One of the main challenges with generative models for periodic materials stems from the creation of representations that are translationally and rotationally invariant. − …”

“…More details of the CDVAE method can be found in Xie et al A recent work by Lyngby and Thygesen successfully applied the CDVAE model to discover new, stable 2D materials and vastly expand the space of 2D materials (on the order of thousands). Another recent work by Moustafa et al used the CDVAE model to discover more than 500 new stable one-dimensional materials. In this work, we trained a CDVAE model (optimizing T c in the latent space) with DFT computed data of 1058 superconducting materials from the JARVIS database and generated thousands of new candidate superconductors.…”

Artificial Intelligence Guided Studies of van der Waals Magnets

“…Competing phases can be investigated by calculating the energy above convex hull. [47,48] Although, metastable phases, that is, phases above the convex hull, may exist in nature. [49,50] We calculated the phonon dispersion for the following promising structures: Cr 2 Fe 2 Br 6 Cl 6 , Fe 2 Cr 2 Br 12 , Mn 2 Cr 2 Cl 12 , and Mn 4 I 6 Cl 6 and found that they were dynamically stable except for Mn 4 I 6 Cl 6 (see Table 2.…”

Design of New Inorganic Crystals with the Desired Composition Using Deep Learning

“…Recent research utilizing this model has successfully achieved the generation of two-dimensional (2D) materials 23 and one-dimensional (1D) materials. 24 These models, however, have difficulty in obtaining the crystals with specific user-desired compositions as they generate crystals with random compositions based on target properties only. The generation and evaluation of materials with the desired compositions play a crucial role in the search for experimentally synthesizable materials.…”

“…With the CDVAE, crystal structures can be searched from a latent space where the information, such as atom types (A), atom coordinates ( X ), and lattice parameters ( L ), are compressed in and generated through the diffusion process. Recent research utilizing this model has successfully achieved the generation of two-dimensional (2D) materials and one-dimensional (1D) materials …”