Information
Registration Number
0222U003257,
0121U108687
,
Title
A theory of phase transformations, kinetic and diffusion phenomena in condensed media
popup.stage_title
Head
Registration Date
22-02-2022
Organization
National Science Center "Kharkiv Institute of Physics and Technology"
popup.description2
1. The behavior of two-layer protective coatings Zn-Mg during annealing is modeled. Annealing promotes the formation of MgZn2 on the upper surface of the coating and Mg2Zn11 on the interface between the Zn layer and the upper layer of ZnMg, which leads to improved corrosion resistance and mechanical characteristics. A diffusion model of this phase transformation is formulated. The basic diffusion equations with moving boundaries are solved numerically applying the transformation of the spatial variable that reduces the diffusion problem to the equivalent diffusion problem with fixed boundaries. The model predictions agree well with experimental results. 2. To study the kinetics of nucleation in condensed binary solutions, a new approach is proposed, which is a combination of classical thermodynamics and macroscopic kinetics. The theory covers the decomposition of liquid and solid solutions proceeding via the nucleation pathway as well as liquid-solid transformation. The cases of nucleation of two-component precipitates of both fixed and variable composition are considered. The equation of equilibrium of the critical nucleus is obtained and its parameters are determined, namely the critical radius and composition. The nucleus equilibrium equations are then used in macroscopic diffusion equations of its growth. This approach can be used in the general case of an non-ideal solution. 3. The problem of non-convex stochastic optimization, usually faced while training deep neural networks, is considered. In physical modeling a similar problem can be solved using the established method of Langevin dynamics in combination with simulated annealing. Based on the solution of the discrete Langevin equation with a slowly increasing viscosity coefficient, a new method of stochastic optimization - CoolMomentum, is proposed. It is shown to achieve high accuracies with the Resnet-20 neural network on the Cifar-10 dataset and the Efficientnet-B0 neural network on the Imagenet dataset.
Product Description
popup.authors
Abyzov Olexander S
Alekseechkin Nikolai V
Bezugly Oleksii I
Borysenko Oleksandr O.
Davydov Leonid M.
Kotlyar Volodymyr V
Lavrova Galyna M.
Lazarev Mykola P.
Mchedlov-Petrosian Petro O.
Poluektov Yurii M
Soroka Oleksii O.
Stepanovsky Yuriy P.
Turkin Anatoliy A.
Chechkin Oleksiy V.
popup.nrat_date
2022-03-09
Updated: 2026-10-13