Leimeroth, Niklas (2024)
Atomistic modelling of crystalline and amorphous Cu-Zr and Si-O-C using machine learning interatomic potentials.
Technische Universität Darmstadt
doi: 10.26083/tuprints-00028306
Dissertation, Erstveröffentlichung, Verlagsversion
Kurzbeschreibung (Abstract)
Modelling and simulation on the atomic scale play a pivotal role for the understanding of complex materials. In this field, machine learning interatomic potentials (MLIPs) are rapidly evolving tools, which allow the description of interatomic interactions with an accuracy approaching that of quantum mechanical methods. At the same time, they are computationally much more efficient, opening the possibility for large-scale molecular dynamics (MD) simulations with unprecedented fidelity. However, current research in the field is often focused on methodical advancements and uses simple single element test cases for this purpose. This thesis treats the development and application of MLIPs, more specifically highly efficient Atomic Cluster Expansion potentials (ACEPs), for structurally and chemically complex systems, namely Cu-Zr and silicon oxycarbide (Si-O-C). Both are representatives of important material classes, metals and glass-ceramics. Cu-Zr has a plethora of intermetallic phases and is a well known metallic glass (MG) former. The performance of the developed potential is compared to previously published classical potentials and experimental data. Using the new MLIP, the concentration-temperature phase diagram of the material is calculated and found to be in good agreement with experiments. Furthermore, the MG structure is investigated, revealing a massively different short-range order compared to classical interatomic potentials (IPs), and tensile tests of a glass-crystal matrix sample show the occurrence of martensitic phase transitions in B2-CuZr. Si-O-C has a highly tunable composition and microstructure. Consequently, training data for this material needs to cover a wide configuration space, which is achieved with an active learning strategy based on structural units present in the bulk material. The developed ACEP is the first publicly available IP for the system and employed to investigate the atomistic structure and its relation to the elastic properties. Contrary to common assumptions, graphite agglomerates in the system are of low importance for the Young’s modulus. Instead, strong correlation to SiO4 tetrahedra and SiC bonds are found. Finally, different types of MLIPs are evaluated. During the work on Cu-Zr and Si-O-C equivariant structure descriptions and message-passing graph neural networks emerged as promising methods to reach ever improving accuracies. Novel NequIP, Allegro and MACE MLIPs implementing them are compared to the well established High-Dimensional Neural Network Potentials (HDNNPs), Gaussian Approximation Potentials (GAPs), Moment Tensor Potentials (MTPs) and ACEPs. The tests reveal the large data requirements for HDNNPs and emphasize the tradeoff between achievable accuracies and computational cost. ACEPs still represent a good compromise in this regard.
Typ des Eintrags: | Dissertation | ||||
---|---|---|---|---|---|
Erschienen: | 2024 | ||||
Autor(en): | Leimeroth, Niklas | ||||
Art des Eintrags: | Erstveröffentlichung | ||||
Titel: | Atomistic modelling of crystalline and amorphous Cu-Zr and Si-O-C using machine learning interatomic potentials | ||||
Sprache: | Englisch | ||||
Referenten: | Albe, Prof. Dr. Karsten ; Drautz, Prof. Dr. Ralf | ||||
Publikationsjahr: | 18 Oktober 2024 | ||||
Ort: | Darmstadt | ||||
Kollation: | xviii, 144 Seiten | ||||
Datum der mündlichen Prüfung: | 2 Oktober 2024 | ||||
DOI: | 10.26083/tuprints-00028306 | ||||
URL / URN: | https://tuprints.ulb.tu-darmstadt.de/28306 | ||||
Kurzbeschreibung (Abstract): | Modelling and simulation on the atomic scale play a pivotal role for the understanding of complex materials. In this field, machine learning interatomic potentials (MLIPs) are rapidly evolving tools, which allow the description of interatomic interactions with an accuracy approaching that of quantum mechanical methods. At the same time, they are computationally much more efficient, opening the possibility for large-scale molecular dynamics (MD) simulations with unprecedented fidelity. However, current research in the field is often focused on methodical advancements and uses simple single element test cases for this purpose. This thesis treats the development and application of MLIPs, more specifically highly efficient Atomic Cluster Expansion potentials (ACEPs), for structurally and chemically complex systems, namely Cu-Zr and silicon oxycarbide (Si-O-C). Both are representatives of important material classes, metals and glass-ceramics. Cu-Zr has a plethora of intermetallic phases and is a well known metallic glass (MG) former. The performance of the developed potential is compared to previously published classical potentials and experimental data. Using the new MLIP, the concentration-temperature phase diagram of the material is calculated and found to be in good agreement with experiments. Furthermore, the MG structure is investigated, revealing a massively different short-range order compared to classical interatomic potentials (IPs), and tensile tests of a glass-crystal matrix sample show the occurrence of martensitic phase transitions in B2-CuZr. Si-O-C has a highly tunable composition and microstructure. Consequently, training data for this material needs to cover a wide configuration space, which is achieved with an active learning strategy based on structural units present in the bulk material. The developed ACEP is the first publicly available IP for the system and employed to investigate the atomistic structure and its relation to the elastic properties. Contrary to common assumptions, graphite agglomerates in the system are of low importance for the Young’s modulus. Instead, strong correlation to SiO4 tetrahedra and SiC bonds are found. Finally, different types of MLIPs are evaluated. During the work on Cu-Zr and Si-O-C equivariant structure descriptions and message-passing graph neural networks emerged as promising methods to reach ever improving accuracies. Novel NequIP, Allegro and MACE MLIPs implementing them are compared to the well established High-Dimensional Neural Network Potentials (HDNNPs), Gaussian Approximation Potentials (GAPs), Moment Tensor Potentials (MTPs) and ACEPs. The tests reveal the large data requirements for HDNNPs and emphasize the tradeoff between achievable accuracies and computational cost. ACEPs still represent a good compromise in this regard. |
||||
Alternatives oder übersetztes Abstract: |
|
||||
Status: | Verlagsversion | ||||
URN: | urn:nbn:de:tuda-tuprints-283064 | ||||
Sachgruppe der Dewey Dezimalklassifikatin (DDC): | 500 Naturwissenschaften und Mathematik > 500 Naturwissenschaften 500 Naturwissenschaften und Mathematik > 530 Physik 500 Naturwissenschaften und Mathematik > 540 Chemie |
||||
Fachbereich(e)/-gebiet(e): | 11 Fachbereich Material- und Geowissenschaften 11 Fachbereich Material- und Geowissenschaften > Materialwissenschaft 11 Fachbereich Material- und Geowissenschaften > Materialwissenschaft > Fachgebiet Materialmodellierung |
||||
Hinterlegungsdatum: | 18 Okt 2024 12:05 | ||||
Letzte Änderung: | 21 Okt 2024 06:03 | ||||
PPN: | |||||
Referenten: | Albe, Prof. Dr. Karsten ; Drautz, Prof. Dr. Ralf | ||||
Datum der mündlichen Prüfung / Verteidigung / mdl. Prüfung: | 2 Oktober 2024 | ||||
Export: | |||||
Suche nach Titel in: | TUfind oder in Google |
Frage zum Eintrag |
Optionen (nur für Redakteure)
Redaktionelle Details anzeigen |