서브메뉴
검색
From Saline to Solids: Studies of Ionic Solvation and Machine Learning for Ab Initio Calculations
From Saline to Solids: Studies of Ionic Solvation and Machine Learning for Ab Initio Calculations
상세정보
- 자료유형
- 학위논문 서양
- 최종처리일시
- 20250211152735
- ISBN
- 9798342729116
- DDC
- 530
- 저자명
- Wills, Alec.
- 서명/저자
- From Saline to Solids: Studies of Ionic Solvation and Machine Learning for Ab Initio Calculations
- 발행사항
- [Sl] : State University of New York at Stony Brook, 2024
- 발행사항
- Ann Arbor : ProQuest Dissertations & Theses, 2024
- 형태사항
- 179 p
- 주기사항
- Source: Dissertations Abstracts International, Volume: 86-05, Section: B.
- 주기사항
- Advisor: Fernandez-Serra, Marivi;Dawber, Matthew.
- 학위논문주기
- Thesis (Ph.D.)--State University of New York at Stony Brook, 2024.
- 초록/해제
- 요약As civilization continues to develop, the challenges faced by the world as a whole grow ever more complex. Climate change has and will continue to drastically alter weather patterns and regional habitability across the globe as energy demands continue to increase. Bacterial resistance to first- and second-line antibiotics, and ever-present threats from quickly mutating viruses can lead to future pandemics that are even harder to manage. Yet, through it all, science must persevere and triumph as we seek solutions to new problems that humanity uncovers. To that end, studies that promote green energy development and biological mechanism characterization are critically needed.Over 97% of the water on our planet contains, to greater or lesser degrees, dissolved salts. It comes as no surprise, then, to find that electrolyte solutions are ubiquitous in nature, ranging from foundational biological structures that permit complex life to new ways for energy storage as we seek to reduce our carbon footprint. It is a natural consequence that an solid understanding of the nature of ion solvation, and the various ion-ion and ion-solvent interactions, is necessary. Through this thorough characterization of solvation, gains in efficiency and efficacy of existing simulation frameworks will allow for a more accurate understanding of the processes at play as salts dissolve, and better inform the development of new methods or materials to advance society as a whole.The nature of solvation is inherently coupled to studies of the electronic interactions between solute and solvent particles. The process of dissolving a salt molecule depends on the interruption of electronic bonds via solvent interactions, which can be modeled in myriad ways. Often, however, simple models serve to inform us about more complicated emergent behaviors in a system. It comes as no surprise, then, that linear response theory can characterize and predict a variety of interesting phenomena, especially concerning electrostatic interactions. Indeed, while many exact descriptions of behaviors are inherently non-linear, a linearization of the problem at hand rarely fails to yield useful insight. Beyond linearizing the solvent response to the solute's presence, simulation frameworks exist which can explore the dynamic properties of the solution as a whole in addition to the pairwise interactions. Classical force fields allow for expedient computation of large system dynamics, while ab initio calculation frameworks allow for accurate electronic structure calculations of those same systems.The work presented here seeks to help inform our understanding of the various attributes commonly seen in free energy landscape of the solvation process. Namely, linear response allows us to inspect interactions to find insight into the behavior of solvation transition states. Further, we help characterize the dependence of the solvation process on the parameter choices one uses when simulating the solution. Finally, it seeks to help determine ways to increase the efficiency of simulating the solvation process by applying machine learning methods to generate model networks that can be used to efficiently simulate more accurate data, in both the classical and ab initio regimes.We introduce the linear response relations that connect the presence of a charge in a dielectric medium to the medium's reaction to that charge. Through these equations, we develop numerical techniques that can be applied to general charge distributions interacting through arbitrary potentials in media with arbitrary dielectric functions. Despite the arbitrariness of these parameters, it is more instructive to focus on the cases of physical relevance. In particular, point charges in dielectric media interacting via the Coulomb potential are given priority. Various approximations for the dielectric function are studied, and various limiting behaviors of the dielectic medium are connected with the behavior of bound states found to exist in these interactions.We then study in detail the solvation process through atomistic simulations of a salt solution -- sodium chloride in water. Through various choices of interaction and model parameter choices, we test the sensitivity of the free energy surface to these parameters in both the classical and ab initio regimes. We show that ab initio parameter choices can have a larger effect on the resulting solution behavior. Further, we apply machine learning paradigms to each of these simulation regimes and study ways to improve the training process in order to better simulate the solvation process. We conclude that representation of the training parameters is an important factor to consider when choosing training data, and that larger training sets do not necessarily mean improved results. When choosing a training subset to maximize representation in the space of training parameters, we find networks with accuracies on par with or better than previously published results. In this vein, we help find ways for others to quickly utilize existing networks and their corresponding training frameworks and be able to extrapolate to new training cases.Finally, we introduce machine learning methods as applied to ab initio simulation frameworks, in particular Kohn-Sham density functional theory. We discuss a number of different frameworks that have been developed for these types of studies. We propose a method for selecting optimally representative subsets of a given training set that shows a strong capability of generating sets that achieve comparable accuracy with an order of magnitude less training data. We then conclude with an introduction of a new framework, titled xcquinox, that promises to extend previously developed packages to include non-locality in the descriptor design, and that is able to easily integrate with new versions of electronic software packages in Python. This interoperability gives us a new, powerful method with which to enhance our learning methods.
- 일반주제명
- Condensed matter physics
- 일반주제명
- Computational physics
- 일반주제명
- Physics
- 일반주제명
- Climate change
- 기타저자
- State University of New York at Stony Brook Physics
- 기본자료저록
- Dissertations Abstracts International. 86-05B.
- 전자적 위치 및 접속
- 로그인 후 원문을 볼 수 있습니다.
MARC
008250123s2024 us c eng d■001000017163648
■00520250211152735
■006m o d
■007cr#unu||||||||
■020 ▼a9798342729116
■035 ▼a(MiAaPQ)AAI31491365
■040 ▼aMiAaPQ▼cMiAaPQ
■0820 ▼a530
■1001 ▼aWills, Alec.
■24510▼aFrom Saline to Solids: Studies of Ionic Solvation and Machine Learning for Ab Initio Calculations
■260 ▼a[Sl]▼bState University of New York at Stony Brook▼c2024
■260 1▼aAnn Arbor▼bProQuest Dissertations & Theses▼c2024
■300 ▼a179 p
■500 ▼aSource: Dissertations Abstracts International, Volume: 86-05, Section: B.
■500 ▼aAdvisor: Fernandez-Serra, Marivi;Dawber, Matthew.
■5021 ▼aThesis (Ph.D.)--State University of New York at Stony Brook, 2024.
■520 ▼aAs civilization continues to develop, the challenges faced by the world as a whole grow ever more complex. Climate change has and will continue to drastically alter weather patterns and regional habitability across the globe as energy demands continue to increase. Bacterial resistance to first- and second-line antibiotics, and ever-present threats from quickly mutating viruses can lead to future pandemics that are even harder to manage. Yet, through it all, science must persevere and triumph as we seek solutions to new problems that humanity uncovers. To that end, studies that promote green energy development and biological mechanism characterization are critically needed.Over 97% of the water on our planet contains, to greater or lesser degrees, dissolved salts. It comes as no surprise, then, to find that electrolyte solutions are ubiquitous in nature, ranging from foundational biological structures that permit complex life to new ways for energy storage as we seek to reduce our carbon footprint. It is a natural consequence that an solid understanding of the nature of ion solvation, and the various ion-ion and ion-solvent interactions, is necessary. Through this thorough characterization of solvation, gains in efficiency and efficacy of existing simulation frameworks will allow for a more accurate understanding of the processes at play as salts dissolve, and better inform the development of new methods or materials to advance society as a whole.The nature of solvation is inherently coupled to studies of the electronic interactions between solute and solvent particles. The process of dissolving a salt molecule depends on the interruption of electronic bonds via solvent interactions, which can be modeled in myriad ways. Often, however, simple models serve to inform us about more complicated emergent behaviors in a system. It comes as no surprise, then, that linear response theory can characterize and predict a variety of interesting phenomena, especially concerning electrostatic interactions. Indeed, while many exact descriptions of behaviors are inherently non-linear, a linearization of the problem at hand rarely fails to yield useful insight. Beyond linearizing the solvent response to the solute's presence, simulation frameworks exist which can explore the dynamic properties of the solution as a whole in addition to the pairwise interactions. Classical force fields allow for expedient computation of large system dynamics, while ab initio calculation frameworks allow for accurate electronic structure calculations of those same systems.The work presented here seeks to help inform our understanding of the various attributes commonly seen in free energy landscape of the solvation process. Namely, linear response allows us to inspect interactions to find insight into the behavior of solvation transition states. Further, we help characterize the dependence of the solvation process on the parameter choices one uses when simulating the solution. Finally, it seeks to help determine ways to increase the efficiency of simulating the solvation process by applying machine learning methods to generate model networks that can be used to efficiently simulate more accurate data, in both the classical and ab initio regimes.We introduce the linear response relations that connect the presence of a charge in a dielectric medium to the medium's reaction to that charge. Through these equations, we develop numerical techniques that can be applied to general charge distributions interacting through arbitrary potentials in media with arbitrary dielectric functions. Despite the arbitrariness of these parameters, it is more instructive to focus on the cases of physical relevance. In particular, point charges in dielectric media interacting via the Coulomb potential are given priority. Various approximations for the dielectric function are studied, and various limiting behaviors of the dielectic medium are connected with the behavior of bound states found to exist in these interactions.We then study in detail the solvation process through atomistic simulations of a salt solution -- sodium chloride in water. Through various choices of interaction and model parameter choices, we test the sensitivity of the free energy surface to these parameters in both the classical and ab initio regimes. We show that ab initio parameter choices can have a larger effect on the resulting solution behavior. Further, we apply machine learning paradigms to each of these simulation regimes and study ways to improve the training process in order to better simulate the solvation process. We conclude that representation of the training parameters is an important factor to consider when choosing training data, and that larger training sets do not necessarily mean improved results. When choosing a training subset to maximize representation in the space of training parameters, we find networks with accuracies on par with or better than previously published results. In this vein, we help find ways for others to quickly utilize existing networks and their corresponding training frameworks and be able to extrapolate to new training cases.Finally, we introduce machine learning methods as applied to ab initio simulation frameworks, in particular Kohn-Sham density functional theory. We discuss a number of different frameworks that have been developed for these types of studies. We propose a method for selecting optimally representative subsets of a given training set that shows a strong capability of generating sets that achieve comparable accuracy with an order of magnitude less training data. We then conclude with an introduction of a new framework, titled xcquinox, that promises to extend previously developed packages to include non-locality in the descriptor design, and that is able to easily integrate with new versions of electronic software packages in Python. This interoperability gives us a new, powerful method with which to enhance our learning methods.
■590 ▼aSchool code: 0771.
■650 4▼aCondensed matter physics
■650 4▼aComputational physics
■650 4▼aPhysics
■650 4▼aClimate change
■653 ▼aSolvent particles
■653 ▼aLinear response theory
■653 ▼aElectrostatic interactions
■690 ▼a0611
■690 ▼a0216
■690 ▼a0605
■690 ▼a0404
■71020▼aState University of New York at Stony Brook▼bPhysics.
■7730 ▼tDissertations Abstracts International▼g86-05B.
■790 ▼a0771
■791 ▼aPh.D.
■792 ▼a2024
■793 ▼aEnglish
■85640▼uhttp://www.riss.kr/pdu/ddodLink.do?id=T17163648▼nKERIS▼z이 자료의 원문은 한국교육학술정보원에서 제공합니다.


