Dokument: Atomic Scale Simulations for Liquid Metals and Alloys: Machine Learning Potentials and Feature Selection
| Titel: | Atomic Scale Simulations for Liquid Metals and Alloys: Machine Learning Potentials and Feature Selection | |||||||
| Weiterer Titel: | Atomic Scale Simulations for Liquid Metals and Alloys: Machine Learning Potentials and Feature Selection | |||||||
| URL für Lesezeichen: | https://docserv.uni-duesseldorf.de/servlets/DocumentServlet?id=71326 | |||||||
| URN (NBN): | urn:nbn:de:hbz:061-20251216-133657-8 | |||||||
| Kollektion: | Dissertationen | |||||||
| Sprache: | Englisch | |||||||
| Dokumententyp: | Wissenschaftliche Abschlussarbeiten » Dissertation | |||||||
| Medientyp: | Text | |||||||
| Autor: | Sandberg, Johannes [Autor] | |||||||
| Dateien: |
| |||||||
| Beitragende: | Prof. Dr. Voigtmann Thomas [Gutachter] Prof. Dr. Horbach, Jürgen [Gutachter] | |||||||
| Dewey Dezimal-Klassifikation: | 500 Naturwissenschaften und Mathematik » 530 Physik | |||||||
| Beschreibungen: | Understanding how atomic interactions give rise to the macroscopic properties of materials is
a main goal of material science. To this end, molecular dynamics simulations are a powerful tool, allowing for directly following the trajectory of atoms in the simulated system, and being well suited for dynamical processes such as diffusion and the early stages of crystallization. This, however, requires the accurate modeling of the interaction between atoms. Empirical interatomic potentials are limited in accuracy, and often fail to reproduce experimental results. Ab initio simulation, while adhering more closely to the underlying quantum-mechanical origin of interactions, is severely limited in scalability, preventing its use in many critical applications. In the past decade, machine-learning potentials, trained on data from ab initio simulation, have become an important method for enabling scalable simulations with ab initio level accuracy. Still, practical ab initio methods rely on approximations, necessitating at some point a connection to be made to experimental results. In this thesis I develop a machine-learned potential for the binary Al-Ni alloy, building upon a previously trained potential for pure Al. For this, the entire process is covered, from construction of the training dataset, to design and training of the potential in the Behler Parrinello high-dimensional neural network framework. The potentials are validated against experimental data for transport coefficients in the liquid state, and applied to the study of homogeneous crystal nucleation from the undercooled liquid, through large-scale molecular dynamics simulation far beyond the reach of ab initio simulation. A significant result of this is the elucidation of the origins of the nucleation pathway into the body-centered cubic B2 phase of equiatomic Al-Ni. I further implement an active feature selection method for such high-dimensional neural network potentials, based on the adaptive group lasso. This allows for reducing the number of input features, taking into account model predictions, allowing for training faster and more explainable potentials. Part of this is the training of a potential for Boron, serving as a particularly complex model system, useful for the evaluation of descriptors and machine-learning potential frameworks.Understanding how atomic interactions give rise to the macroscopic properties of materials is a main goal of material science. To this end, molecular dynamics simulations are a powerful tool, allowing for directly following the trajectory of atoms in the simulated system, and being well suited for dynamical processes such as diffusion and the early stages of crystallization. This, however, requires the accurate modeling of the interaction between atoms. Empirical interatomic potentials are limited in accuracy, and often fail to reproduce experimental results. Ab initio simulation, while adhering more closely to the underlying quantum-mechanical origin of interactions, is severely limited in scalability, preventing its use in many critical applications. In the past decade, machine-learning potentials, trained on data from ab initio simulation, have become an important method for enabling scalable simulations with ab initio level accuracy. Still, practical ab initio methods rely on approximations, necessitating at some point a connection to be made to experimental results. In this thesis I develop a machine-learned potential for the binary Al-Ni alloy, building upon a previously trained potential for pure Al. For this, the entire process is covered, from construction of the training dataset, to design and training of the potential in the Behler Parrinello high-dimensional neural network framework. The potentials are validated against experimental data for transport coefficients in the liquid state, and applied to the study of homogeneous crystal nucleation from the undercooled liquid, through large-scale molecular dynamics simulation far beyond the reach of ab initio simulation. A significant result of this is the elucidation of the origins of the nucleation pathway into the body-centered cubic B2 phase of equiatomic Al-Ni. I further implement an active feature selection method for such high-dimensional neural network potentials, based on the adaptive group lasso. This allows for reducing the number of input features, taking into account model predictions, allowing for training faster and more explainable potentials. Part of this is the training of a potential for Boron, serving as a particularly complex model system, useful for the evaluation of descriptors and machine-learning potential frameworks. | |||||||
| Lizenz: |  Dieses Werk ist lizenziert unter einer Creative Commons Namensnennung 4.0 International Lizenz | |||||||
| Fachbereich / Einrichtung: | Mathematisch- Naturwissenschaftliche Fakultät | |||||||
| Dokument erstellt am: | 16.12.2025 | |||||||
| Dateien geändert am: | 17.12.2025 | |||||||
| Promotionsantrag am: | 07.10.2024 | |||||||
| Datum der Promotion: | 17.12.2024 |
