Speaker
Description
I will present a methodology of constructing phase diagrams by machine-learning the free energy function from data resulting from molecular simulations (such as lattice dynamics, phase coexistence simulations, or phonon calculations). In a nutshell, the methodology can be described as thermodynamic integration, but with active sampling and uncertainty estimation capabilities. This methodology is coupled with Moment Tensor Potentials, a class of machine-learning potentials capable of actively learning the underlying quantum-mechanical potential energy surface. Thus, the combined methodology allows for automatic construction of phase diagrams given a set of phases and a quantum-mechanical solver, where machine-learning algorithms provide the bridge between the scales.
The methodology will be illustrated on two examples: constructing the Lennard-Jones phase diagram which is excellent for benchmarking, and a phase diagram of lithium at moderate pressures and temperatures.
This work is supported by the Russian Science Foundation, grant number 18-13-00479.
| Speaker Country | Russia |
|---|