Predicting materials' properties with combining machine learning and classical molecular dynamical simulation.

Creating materials thermodynamical features (such as heat capacity) by MD simulation (LAMMPS) only from chemical formulae, where the interatomic potential is automatically selected from predefined by atom-atom combination. Those generated features are then plugged into the machine learning estimator (here lightGBM is used) to predict materials' properties. (https://github.com/LazyDragon1123/Materials_AI)

Motivated by the graphical representation of crystals as feature vectors, originally introduced in CGCNN by Tian Xie (https://arxiv.org/abs/1710.10324), we create 'magnetic graph representation', where magnetic sites and the spin-spin exchange interaction is incorporated into the node and edge in the graph, respectively. This representation enhances the original CGCNN's predictive performances. (https://github.com/LazyDragon1123/Magnetic_CGCNN)

Classification of Ferromagnetic ordering Antiferromagnetic ordering is tested for both MCGCCNN and CGCNN, where the prediction's performance is slightly improved as shown below.