1-2、Machine-learned O(N) interatomic potentials for DFT-accurate large-scale simulations of amorphous materials

Stephen Elliott

University of Cambridge

摘要：The most accurate computer simulations of materials are currently carried out using density-functional theory (DFT)-based codes, in which the electronic interactions are modelled using quantum-mechanical calculations, albeit with certain necessary approximations in order to make the calculations tractable. However, the severe drawback with such an approach is that it is computationally extremely demanding, since the computational time scales as O(N3), cubically with the number of atoms (N) in the model. This means that it is not possible to simulate models containing more than a few hundred atoms, for simulation times of less than a nanosecond or so, even using the most powerful available supercomputers. While such simulations are sufficient to model local properties of materials, they are obviously insufficient to capture more long-range physical behaviour, e.g. mechanical or thermal properties. This is particularly problematic and acute for the case of modelling non-crystalline (amorphous/ glassy) materials for which there is no unit cell (effectively of infinite extent) due to the lack of long-range translational periodicity. Although a large number of empirical interatomic potentials are available for a very wide range of material systems, invariably these potentials are not electronically accurate. There is, therefore, a pressing need to develop interatomic potentials that are linear-scaling (O(N)) but which nevertheless retain a DFT level of accuracy. The computational speed of these O(N) potentials can then allow the construction of an ensemble of models for a particular material composition, thereby permitting statistically significant computed results to be obtained for the models, unachievable for the usually single models reported from conventional DFT-based molecular-dynamics simulations.

We have recently developed a number of O(N) interatomic (‘Gaussian Approximation’) potentials (GAPs) for monatomic materials (C, Si), binary materials (C:Li,Na) and ternary materials (Ge-Sb-Te), using a machine-learning approach to fit electronic-energy surfaces for a wide range of DFT-constructed atomic configurations. We have used these potentials to simulate models of: (i) amorphous (a-) C and C:Li,Na with varying levels of porosity, in connection with battery-cathode applications; (ii) models of a-Si with varying levels of structural order resulting from increasingly slow quench rates, for potential low-mechanical-loss applications in the Bragg mirrors of the LIGO interferometer; and (iii) models of a-Ge2Sb2Te5 for non-volatile phase-change random-access memory applications.

DOI:10.12110/secondfmge.20181014.102