Investigating Non-Periodic Solids Using First Principles Calculations and Machine Learning Algorithms

DSpace/Manakin Repository

Investigating Non-Periodic Solids Using First Principles Calculations and Machine Learning Algorithms

Citable link to this page

 

 
Title: Investigating Non-Periodic Solids Using First Principles Calculations and Machine Learning Algorithms
Author: Çubuk, Ekin Doğuş
Citation: Çubuk, Ekin Doğuş. 2016. Investigating Non-Periodic Solids Using First Principles Calculations and Machine Learning Algorithms. Doctoral dissertation, Harvard University, Graduate School of Arts & Sciences.
Full Text & Related Files:
Abstract: Computational methods are expected to play an increasingly important role in materials design. In order to live up to these expectations, simulations need to have predictive power. To achieve this, there are two hurdles, both relating to the complexity of physical interactions. The first is the quantum mechanical interactions of ions and electrons at short timescales, which have proven difficult to simulate using classical computation. While it is now possible to model some properties and materials using first principles methods (e.g. density functional theory), accuracy, consistency and computational efficiency need to be improved to meet the demands of high-throughput materials design. The second hurdle is the difficulty of predicting the outcomes of interactions between ions at longer timescales. These interactions are central to some of the biggest mysteries in condensed matter physics, such as the glass transition.

Meanwhile, the field of machine learning and artificial intelligence has seen rapid progress in the last decade. Due to improvements in hardware, software, and methodology, machine learning algorithms are now able to learn complex tasks by mastering fundamental concepts from data.

Thus, this thesis explores the applicability of machine learning to the main challenges facing computational materials design. First, as a case study, we investigate the lithiation of amorphous silicon. We show that large unit cells need to be simulated to model lithium-silicon alloys accurately. By analyzing the geometric structures of local neighborhoods of silicon atoms, it is possible to explain the macroscopic behavior from microscopic signatures. In response to the first hurdle as discussed above, we train neural networks to reproduce energies of silicon structures and silicon-lithium alloys, which allows us to study much larger unit cells. We then explore silicon neural networks in detail, in order to explain how this specific machine learning architecture can model quantum mechanical interactions.

The following two chapters focus on the second hurdle which arises from complex ionic configurations. By studying Lennard-Jones supercooled liquids, we try to resolve two mysteries related to supercooled liquids: 1) why the dynamics are spatially heterogeneous, and 2) why the relaxation time increases super-exponentially as the temperature is lowered. Through machine learning, we can resolve the first mystery quantitatively. Furthermore, we show that the second can also be resolved in our framework, by using empirical measurements of the machine learned representation, which we call ``softness''. Finally, we discuss the physical meaning of softness, by comparing it to other measures and applying unsupervised learning and reduced curve-fitting models.
Terms of Use: This article is made available under the terms and conditions applicable to Other Posted Material, as set forth at http://nrs.harvard.edu/urn-3:HUL.InstRepos:dash.current.terms-of-use#LAA
Citable link to this page: http://nrs.harvard.edu/urn-3:HUL.InstRepos:33493370
Downloads of this work:

Show full Dublin Core record

This item appears in the following Collection(s)

 
 

Search DASH


Advanced Search
 
 

Submitters