GN-OA is a crystal structure prediction tool, which can predict crystal structures from scratch with extremely low computational cost.
In GN-OA a graph network (GN) is adopted to establish a correlation model between the crystal structure and formation enthalpies, and optimization algorithm (OA) is used to accelerate the search for crystal structure with optimal formation enthalpy. The approach of combining GN and OA for crystal structure searching (GN-OA) can predict crystal structures at given chemical compositions with additional constraints on cell shapes and lattice symmetries.
(1) Data preparation
a) MatBench dataset
See the file
If you want to use OQMD, you need to download the database first. Download
(2) Training the GN model
We provided the origin MEGNet model in
./NN_model/orig_megnet.py, you just need to run this python file to get a trained GN model.
Of course, you can also build a new GN model.
(1) Input file
[BASE] # The chemical formula of the compound, element symbol + count, i.e., Ca4 S4, Cs1 Pb1 I3 compound = Ca4 S4 # The GN model file path, it is better to use absolute path. gn_model_path = /GN/model/path # Output path, use to save the results. output_path = . # Load model and predict using GPU use_gpu = False [LATTICE] # The range of space group: # [A, B]-> A<=sg<=B space_group = [2, 230] # Lattice a,b,c (unit: Angstrom): # [A, B] -> A<=Lattice<=B lattice_a = [2, 30] lattice_b = [2, 30] lattice_c = [2, 30] # Lattice alpha,beta,gamma (unit: degree): # [A, B] -> A<=Lattice<=B lattice_alpha = [20, 160] lattice_beta = [20, 160] lattice_gamma = [20, 160] [PROGRAM] # Search algorithm: # 1) 'tpe' (Bayesian Optimization); # 2) 'pso' (Particle Swarm Optimization) # 3) 'rand' (Random Search); algorithm = tpe # 1) If `algorithm = tpe`, `n_init` is the number of initial random points; # 2) If `algorithm = pso`, `n_init` is the number of particle; # 3) If `algorithm = rand`, `n_init` is invalid. n_init = 200 # The maximum steps of program runs # 1) If `algorithm = tpe`, `max_step` is the maximum steps of program runs; # 2) If `algorithm = pso`, `max_step` is the maximum iterations; # 3) If `algorithm = rand`, `max_step` is the maximum steps of program runs. max_step = 5000 # Specify the random seed, -1 is None rand_seed = 100
(2) Search crystal structure
predict_structure.py file in the path where the input file
cd [/input/file/path] python [/the/project/path]/predict_structure.py
If you use GN-OA for research, please consider citing this paper: Guanjian Cheng, Xin-Gao Gong, Wan-Jian Yin. “Crystal structure prediction by combining graph network and optimization algorithm.” Nat Commun 2022;13(1):1492. Doi: 10.1038/s41467-022-29241-4.