서브메뉴
검색
Machine Learning Methods for Cross Section Measurements
Machine Learning Methods for Cross Section Measurements
상세정보
- 자료유형
- 학위논문 서양
- 최종처리일시
- 20260202105110
- ISBN
- 9798293893348
- DDC
- 530
- 저자명
- Desai, Krish.
- 서명/저자
- Machine Learning Methods for Cross Section Measurements
- 발행사항
- [Sl] : University of California, Berkeley, 2025
- 발행사항
- Ann Arbor : ProQuest Dissertations & Theses, 2025
- 형태사항
- 828 p
- 주기사항
- Source: Dissertations Abstracts International, Volume: 87-04, Section: B.
- 주기사항
- Advisor: Nachman, Benjamin;Seljak, Uros.
- 학위논문주기
- Thesis (Ph.D.)--University of California, Berkeley, 2025.
- 초록/해제
- 요약Precise differential cross section measurements are indispensable for tests of Standard Model predictions at the energy frontier and for searches for new physics, yet their extraction from collider data is an ill posed inverse problem. Unfolding, also known as deconvolution, is the process of removing detector distortions to reconstruct particle level truth from detector level data. Conventional, histogram based, binned unfolding techniques introduce artifacts, impose arbitrary bin edges, and become computationally prohibitive in high dimensional phase spaces, potentially obscuring underlying physics.This dissertation develops a unified framework that leverages modern machine learning techniques to surmount these limitations. First, the Neural Posterior Unfolding (NPU) method demonstrates how conditional normalising flows can serve as differentiable surrogates of detector response, enabling likelihood based unfolding through implicit regularisation. Building on this foundation, the Moment Unfolding algorithm directly extracts distribution moments without binning, providing precise experimental predictions for effective field theories and phenomenological models. The framework is further advanced by development of Reweighting Adversarial Networks (RANs), which perform full spectral unfolding using adversarial training to implement particle level reweighting guided by a detector level classifier, offering theoretical and computational advantages over extant methods. A critical statistical analysis of event correlations in unfolded data reveals systematic misestimation of uncertainties when these correlations are ignored, leading to methodological recommendations that ensure correct coverage for all derived observables. The methods presented in this dissertation are validated using both idealised Gaussian distributions and proton--proton collision simulations of the CMS experiment as realistic particle physics examples, specifically simulations of Z+jets events, demonstrating significant improvements in precision, accuracy, and computational efficiency, reducing computational time by orders of magnitude while maintaining or exceeding the precision of existing methods for the unbiased recovery of spectral features. By marrying statistical rigour with powerful machine learning methods, this work establishes a scalable blueprint for precision measurements at current and future high energy physics experiments. The resulting open source software enables more reliable extraction of fundamental physics parameters from complex detector data, advancing the ability to test theoretical models and potentially discover new phenomena.
- 일반주제명
- Physics
- 일반주제명
- Particle physics
- 일반주제명
- Computational physics
- 키워드
- Cross sections
- 키워드
- Deconvolution
- 키워드
- Machine learning
- 키워드
- Unfolding
- 기타저자
- University of California, Berkeley Physics
- 기본자료저록
- Dissertations Abstracts International. 87-04B.
- 전자적 위치 및 접속
- 로그인 후 원문을 볼 수 있습니다.
MARC
008260126s2025 us c eng d■001000017359375
■00520260202105110
■006m o d
■007cr#unu||||||||
■020 ▼a9798293893348
■035 ▼a(MiAaPQ)AAI32237024
■040 ▼aMiAaPQ▼cMiAaPQ
■0820 ▼a530
■1001 ▼aDesai, Krish.
■24510▼aMachine Learning Methods for Cross Section Measurements
■260 ▼a[Sl]▼bUniversity of California, Berkeley▼c2025
■260 1▼aAnn Arbor▼bProQuest Dissertations & Theses▼c2025
■300 ▼a828 p
■500 ▼aSource: Dissertations Abstracts International, Volume: 87-04, Section: B.
■500 ▼aAdvisor: Nachman, Benjamin;Seljak, Uros.
■5021 ▼aThesis (Ph.D.)--University of California, Berkeley, 2025.
■520 ▼aPrecise differential cross section measurements are indispensable for tests of Standard Model predictions at the energy frontier and for searches for new physics, yet their extraction from collider data is an ill posed inverse problem. Unfolding, also known as deconvolution, is the process of removing detector distortions to reconstruct particle level truth from detector level data. Conventional, histogram based, binned unfolding techniques introduce artifacts, impose arbitrary bin edges, and become computationally prohibitive in high dimensional phase spaces, potentially obscuring underlying physics.This dissertation develops a unified framework that leverages modern machine learning techniques to surmount these limitations. First, the Neural Posterior Unfolding (NPU) method demonstrates how conditional normalising flows can serve as differentiable surrogates of detector response, enabling likelihood based unfolding through implicit regularisation. Building on this foundation, the Moment Unfolding algorithm directly extracts distribution moments without binning, providing precise experimental predictions for effective field theories and phenomenological models. The framework is further advanced by development of Reweighting Adversarial Networks (RANs), which perform full spectral unfolding using adversarial training to implement particle level reweighting guided by a detector level classifier, offering theoretical and computational advantages over extant methods. A critical statistical analysis of event correlations in unfolded data reveals systematic misestimation of uncertainties when these correlations are ignored, leading to methodological recommendations that ensure correct coverage for all derived observables. The methods presented in this dissertation are validated using both idealised Gaussian distributions and proton--proton collision simulations of the CMS experiment as realistic particle physics examples, specifically simulations of Z+jets events, demonstrating significant improvements in precision, accuracy, and computational efficiency, reducing computational time by orders of magnitude while maintaining or exceeding the precision of existing methods for the unbiased recovery of spectral features. By marrying statistical rigour with powerful machine learning methods, this work establishes a scalable blueprint for precision measurements at current and future high energy physics experiments. The resulting open source software enables more reliable extraction of fundamental physics parameters from complex detector data, advancing the ability to test theoretical models and potentially discover new phenomena.
■590 ▼aSchool code: 0028.
■650 4▼aPhysics
■650 4▼aParticle physics
■650 4▼aComputational physics
■653 ▼aCross sections
■653 ▼aDeconvolution
■653 ▼aHigh energy physics
■653 ▼aMachine learning
■653 ▼aUnfolding
■690 ▼a0605
■690 ▼a0798
■690 ▼a0216
■71020▼aUniversity of California, Berkeley▼bPhysics.
■7730 ▼tDissertations Abstracts International▼g87-04B.
■790 ▼a0028
■791 ▼aPh.D.
■792 ▼a2025
■793 ▼aEnglish
■85640▼uhttp://www.riss.kr/pdu/ddodLink.do?id=T17359375▼nKERIS▼z이 자료의 원문은 한국교육학술정보원에서 제공합니다.
