-
Gradient-Based Training and Pruning of Radial Basis Function Networks with an Application in Materials Physics
Authors:
Jussi Määttä,
Viacheslav Bazaliy,
Jyri Kimari,
Flyura Djurabekova,
Kai Nordlund,
Teemu Roos
Abstract:
Many applications, especially in physics and other sciences, call for easily interpretable and robust machine learning techniques. We propose a fully gradient-based technique for training radial basis function networks with an efficient and scalable open-source implementation. We derive novel closed-form optimization criteria for pruning the models for continuous as well as binary data which arise…
▽ More
Many applications, especially in physics and other sciences, call for easily interpretable and robust machine learning techniques. We propose a fully gradient-based technique for training radial basis function networks with an efficient and scalable open-source implementation. We derive novel closed-form optimization criteria for pruning the models for continuous as well as binary data which arise in a challenging real-world material physics problem. The pruned models are optimized to provide compact and interpretable versions of larger models based on informed assumptions about the data distribution. Visualizations of the pruned models provide insight into the atomic configurations that determine atom-level migration processes in solid matter; these results may inform future research on designing more suitable descriptors for use with machine learning algorithms.
△ Less
Submitted 6 April, 2020;
originally announced April 2020.
-
Dynamic coupling between particle-in-cell and atomistic simulations
Authors:
Mihkel Veske,
Andreas Kyritsakis,
Kyrre Ness Sjobak,
Vahur Zadin,
Alvo Aabloo,
Flyura Djurabekova
Abstract:
We propose a method to directly couple molecular dynamics, finite element method and particle-in-cell techniques to simulate metal surface response to high electric fields. We use this method to simulate the evolution of a field emitting tip under thermal runaway by fully including the 3D space-charge effects. We also present a comparison of the runaway process between the two tip geometries of di…
▽ More
We propose a method to directly couple molecular dynamics, finite element method and particle-in-cell techniques to simulate metal surface response to high electric fields. We use this method to simulate the evolution of a field emitting tip under thermal runaway by fully including the 3D space-charge effects. We also present a comparison of the runaway process between the two tip geometries of different widths. The results show with high statistical significance that in case of sufficiently narrow field emitters, the thermal runaway occurs in cycles where intensive neutral evaporation alternates with the cooling periods. The comparison with previous works shows, that the evaporation rate in the regime of intensive evaporation is sufficient to ignite a plasma arc above the simulated field emitters.
△ Less
Submitted 24 May, 2020; v1 submitted 19 June, 2019;
originally announced June 2019.
-
Dynamic coupling of a finite element solver to large-scale atomistic simulations
Authors:
Mihkel Veske,
Andreas Kyritsakis,
Kristjan Eimre,
Vahur Zadin,
Alvo Aabloo,
Flyura Djurabekova
Abstract:
We propose a method for efficiently coupling the finite element method with atomistic simulations, while using molecular dynamics or kinetic Monte Carlo techniques. Our method can dynamically build an optimized unstructured mesh that follows the geometry defined by atomistic data. On this mesh, different multiphysics problems can be solved to obtain distributions of physical quantities of interest…
▽ More
We propose a method for efficiently coupling the finite element method with atomistic simulations, while using molecular dynamics or kinetic Monte Carlo techniques. Our method can dynamically build an optimized unstructured mesh that follows the geometry defined by atomistic data. On this mesh, different multiphysics problems can be solved to obtain distributions of physical quantities of interest, which can be fed back to the atomistic system. The simulation flow is optimized to maximize computational efficiency while maintaining good accuracy. This is achieved by providing the modules for a) optimization of the density of the generated mesh according to requirements of a specific geometry and b) efficient extension of the finite element domain without a need to extend the atomistic one. Our method is organized as an open-source C++ code. In the current implementation, an efficient Laplace equation solver for calculation of electric field distribution near rough atomistic surface demonstrates the capability of the suggested approach.
△ Less
Submitted 19 January, 2018; v1 submitted 29 June, 2017;
originally announced June 2017.
