MAGUS: machine learning and graph theory assisted universal structure searcher

Abstract

Crystal structure predictions based on first-principles calculations have gained great success in materials science and solid state physics. However, the remaining challenges still limit their applications in systems with a large number of atoms, especially the complexity of conformational space and the cost of local optimizations for big systems. Here, we introduce a crystal structure prediction method MAGUS based on the evolutionary algorithm, which addresses the above challenges with machine learning and graph theory. Techniques used in the program are summarized in detail and benchmark tests are provided. With intensive tests, we demonstrate that on-the-fly machine learning potentials can be used to reduce the number of expensive first-principles calculations significantly, and the crystal decomposition based on graph theory can efficiently decrease the required configurations to find the target structures. We also summarized the representative applications of this method on several research topics, including unexpected compounds in the interior of planets and their exotic states at high pressure and high temperature (superionic, plastic, partially diffusive state, etc.); new functional materials (superhard, high-energy-density, superconducting, photoelectric materials), etc. These successful applications demonstrated that MAGUS code can help to accelerate the discovery of interesting materials and surprising physical phenomena, as well as the significant value of crystal structure predictions in general.

Graph theory and graph neural network assisted high-throughput crystal structure prediction and screening for energy conversion and storage

“…Therefore, we believe that the CHGNet MLP has decent accuracy in predicting dynamic stability. The 61.5% success rate of positive dispersions as validated by DFT calculations is good enough for high-throughput screening of a large number of previously unexplored structures if the initial screening pool is very large (say >10 4 predicted positive dispersions). The DFT validation result conrms the high success rate of using the pre-trained CHGNet MLP for quickly screening dynamic stability, which would have taken much longer simulation time and resources if running by full DFT calculations.…”

“…In this paper, we established a workflow consisting of hypothetical structure generation, fast structure optimization, and quick screening of dynamical stability for searching for ultralow LTC for thermoelectric energy conversion and high heat capacity for thermal energy storage. We first use an ML algorithm called “machine learning and graph theory assisted universal structure searcher (MAGUS) 4 ” for the generation of large-scale hypothetical structures, followed by fast structure optimization and quick screening of dynamical stability using a pre-trained MLP named “Crystal Hamiltonian Graph neural Network (CHGNet)”. 28 The LTC and heat capacity of the filtered structures in the previous steps are predicted by our separately trained graph neural network (GNN) models.…”

“…Recently, the emergence of computational methods has offered an alternative avenue, thanks to the advancement of computational soware and the availability of supercomputing resources. 4,5 This high-throughput computation of crystal structure prediction (CSP) at the quantum level, although accurate to the largest extent, is also time and computer resource consuming. Hence, to date, it is still computationally expensive to search for large-scale new and stable structures, despite some success on a few specic material families or small number of systems.…”

Strategies for improving the superconductivity of hydrides under high pressure

“…Currently, research on hydrogen-based superconductors is primarily driven by computation-led experiments. Through utilizing advanced evolutionary/genetic algorithms [57][58][59][60], random structure searching [61,62], particle swarm optimization [63,64], minima hopping [65,66], metadynamics [67], simulated annealing [68,69] machine learning and graph theory assisted universal structure searcher method [70] and crystal structure analysis by evolutional local random computational method [71] in conjunction with first-principles calculations, the ground-state structure can be determined solely based on the given chemical composition and pressures. To provide readers with guidance on effectively utilizing these software tools for structural prediction, we have selected three popular structural prediction software tools, USPEX, CALYPSO and AIRSS.…”

Theoretical Predictions on Superconducting Phase above Room Temperature in Lutetium-Beryllium Hydrides at High Pressures

“…The results are verified by the graph theory assisted universal structure searcher MAGUS. [31] We then re-optimized the structures using the ab initio calculation of the Quantum Espresso (QE) package, [32] and calculated the phonon spectrum at different pressures using the density functional perturbation theory (DFPT). [33] The charge density and the wave function cutoff values are 600 Ry and 60 Ry, respectively.…”