Long, Teng (2022)
Accelerating the discovery of crystalline materials with desired intrinsic properties by machine learning.
Technische Universität Darmstadt
doi: 10.26083/tuprints-00019964
Ph.D. Thesis, Primary publication, Publisher's Version
|
Text
Accelerating_the_discovery_of_crystalline_materials_with_desired_intrinsic_properties_by_machine_learning.pdf Copyright Information: CC BY 4.0 International - Creative Commons, Attribution. Download (13MB) | Preview |
Item Type: | Ph.D. Thesis | ||||
---|---|---|---|---|---|
Type of entry: | Primary publication | ||||
Title: | Accelerating the discovery of crystalline materials with desired intrinsic properties by machine learning | ||||
Language: | English | ||||
Referees: | Zhang, Prof. Dr. Hongbin ; Xu, Prof. Dr. Baixiang | ||||
Date: | 2022 | ||||
Place of Publication: | Darmstadt | ||||
Collation: | xxi, 180 Seiten | ||||
Date of oral examination: | 11 November 2021 | ||||
DOI: | 10.26083/tuprints-00019964 | ||||
Abstract: | As an emergent research paradigm, data-driven methods (e.g., machine learning) have recently been applied extensively to materials science, which provide valuable solutions to map out the process-structure-property relationships, thus enabling autonomous materials designs. In this thesis, focusing on the mapping between crystal structures and intrinsic physical properties, both forward modelling (to predict physical properties with crystal structures as input) and inverse design (to predict novel crystal structures with desired properties) have been performed, accelerating the design of crystalline materials with desired properties. For the forward modelling, Curie temperature of 1749 ferromagnetic materials was collected to carry out machine learning modelling based on the two-step random forest method. The resulting accuracy is about 91% for evaluating the Curie temperature, which has been further validated by 85 experimental results. In this regard, it provides a practical solution to accelerate designing functional ferromagnetic materials, as the Curie temperature is one of the three key intrinsic magnetic properties (in addition to magnetization and magnetic anisotropy energy). Furthermore, in collaboration with Yixuan Zhang, we demonstrated that both the total energies and forces on atoms could be modelled accurately, leading to a reliable construction of machine-learning interatomic potentials for further atomistic simulations like molecular dynamics. Therefore, the forward modelling can be applied to predict the intrinsic physical properties and to bridge quantitative simulations across the electronic and atomistic length scales. In terms of inverse design, constrained crystal deep convolutional generative adversarial networks (CCDCGAN) have been developed, directly predicting crystal structures distinct from the known cases based on the image-based continuous representation (of the crystal structures) forming a latent space. Moreover, the intrinsic properties of generated structures can be optimized in the latent space through a back propagator (applied on the pre-trained forward model) and an appropriately defined loss function (of CCDCGAN). In this thesis, formation energy has been integrated into CCDCGAN as a forward model, and correspondingly CCDCGAN can design stable crystal structures. It has been successfully applied on a binary system (Bi-Se) and multicomponent systems (binary, ternary, and quaternary systems). It is observed that unreported crystal structures below the convex hull defined by the known experimental cases can be obtained. Interestingly, it is suspected that the CCDCGAN model can be generalized to multi-objective optimization when forward models on different properties are applied, leading to a systematic way of designing novel crystalline materials. In the future, inverse design of crystalline materials with multi-objective optimization of various physical properties can be realized based on our current implementation of CCDCGAN, e.g., by incorporating mechanical and other magnetic properties to design magnetic materials with optimal performance. Additionally, such image-based deep learning algorithms can be straightforwardly generalized to model the microstructures, which will provide a systematic method to inverse design the microstructures and hence a decent solution to map out the structure-property relationship. |
||||
Alternative Abstract: |
|
||||
Status: | Publisher's Version | ||||
URN: | urn:nbn:de:tuda-tuprints-199642 | ||||
Classification DDC: | 500 Science and mathematics > 500 Science | ||||
Divisions: | 11 Department of Materials and Earth Sciences > Material Science | ||||
Date Deposited: | 06 Jan 2022 13:22 | ||||
Last Modified: | 06 Jan 2022 13:22 | ||||
URI: | https://tuprints.ulb.tu-darmstadt.de/id/eprint/19964 | ||||
PPN: | 490509363 | ||||
Export: |
View Item |