Elements: Portable Machine Learning Models for Experimental Nuclear Physics

核物理实验产生越来越多的大量数据，需要新颖和复杂的数据分析方法来支持科学发现的过程。机器学习（ML）领域位于计算机科学、统计学和数学的交叉点，它关注的是开发算法和系统，这些算法和系统可以通过积累的经验来改进特定的任务。与依赖程序员明确指令的传统软件系统相比，机器学习系统能够自动从大型数据集中获得见解。这种方法对许多学科产生了重大影响——事实上，几乎影响了社会的每个领域——核物理学也不例外。然而，当一个人调查了物理社区中使用ML的无数种方式时，一个显著的趋势出现了，即为每个新应用从头开始构建定制模型的趋势。这是一项需要大量技术专长和大量数据的任务，这些数据是由人类精心注释的，统计模型可以从中“学习”。机器学习的最新进展集中在最大限度地减少对大规模手工标记数据的依赖。这个项目的核心是利用这些新技术在密歇根州的稀有同位素束设施（FRIB）建立各种核物理实验模型，然后将其发布给科学界。最终用户可以通过流程离线调优来调整这些模型，以适应各种下游应用程序。这些模型是使用来自FRIB三个粒子探测器系统的未标记数据开发的-主动目标时间投影室（at - tpc）， Summing NaI （SuN）探测器和SAMURAI Pion重建和离子跟踪器（SPiRIT）时间投影室。我们在FRIB的合作者已经确定了关键分析和拟合任务，对这些模型进行了评估。本科生在执行研究议程方面发挥着核心作用。通过与国家机构的物理学家合作，让他们参与尖端研究，学生们可以为在学术界和更广泛的劳动力中从事有影响力的STEM职业做好准备。此外，该项目与戴维森学院的科学体验研究项目合作，为历史上被排除在科学之外的群体的学生提供我们实验室的整个夏季研究经验。使用最先进的自监督机器学习（ML）方法构建预训练模型，以支持希望解决核物理实验中各种分析任务的物理学家。这些模型支持三种探测器系统的用户，这些系统是由密歇根州稀有同位素光束设施（FRIB）的科学家小组开发和维护的。FRIB是一个核科学用户设施，于2022年夏天上线。FRIB的用户在整个核景观中调查核性质，全年需要各种不同的实验装置。由于给定探测器中实验之间的光束组成和能量变化等参数，设备用户需要为每个实验重新训练新的ML模型。预训练模型可以快速调整和适应用户期望的用例，减少实验人员的负担。这项工作建立在行业标准和行业领先的机器学习软件和库（如pytorch和tensorflow）的基础上，预训练的模型对社区开放，以及用于针对特定应用进行微调的脚本。这些模型也被集成到领域科学家使用的标准软件中。这项工作汇集了实验核物理学家、计算机科学家和数据科学家，创造了一个跨学科的紧密合作。本科生在执行研究议程方面发挥着核心作用。通过与国家机构的物理学家合作，让他们参与尖端研究，学生们可以为在学术界和更广泛的劳动力中从事有影响力的STEM职业做好准备。该奖项由先进网络基础设施办公室颁发，由数学和物理科学理事会物理部的信息前沿物理项目联合支持。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

Experiments in nuclear physics produce increasingly vast amounts of data that require novel and sophisticated methods of data analysis to support the process of scientific discovery. The field of machine learning (ML), which sits at the intersection of computer science, statistics and mathematics, is concerned with developing algorithms and systems that improve at a certain task with accrued experience. In contrast to traditional software systems that rely on explicit instructions from the programmer, ML systems are able to automatically derive insights from large datasets. This approach has had a significant impact on many disciplines—indeed, on nearly every area of society—and nuclear physics is no exception. However, when one surveys the myriad ways in which ML is used in the physics community, a notable trend emerges, namely, the tendency to construct bespoke models from scratch for each new application. This is a task that requires considerable technical expertise and access to large volumes of data that have been painstakingly annotated by humans from which the statistical models can “learn”. Recent advances in ML have focused on minimizing the reliance on large-scale hand-labeling of data. This project centers on building models using these new techniques for various nuclear physics experiments at the Facility for Rare Isotope Beams (FRIB) in Michigan that will then be released to the scientific community. These models can be adapted through a process offine-tuning by end-users for a variety of downstream applications. The models are developed using unlabeled data from three particle detector systems at FRIB – the Active-Target Time Projection Chamber (AT-TPC), the Summing NaI (SuN) detector, and the SAMURAI Pion Reconstruction and Ion Tracker (SPiRIT) Time Projection Chamber. The models are evaluated on key analysis and fitting tasks that have been identified by our collaborators at FRIB. Undergraduate students play a central role in executing the research agenda. By engaging them in cutting-edge research in partnership with physicists at a national facility, students are prepared for impactful careers in STEM, both in academia and in the broader workforce. Additionally, this project has partnered with the Research in Science Experience program at Davidson College and provides full-summer research experiences in our lab to students from groups historically excluded from the sciences.Pretrained models are built using state-of-the-art self-supervised machine learning (ML) methods to support physicists who would like to solve a variety of analysis tasks in nuclear physics experiments. These models support users of three detector systems that are developed and maintained by groups of scientists at the Facility for Rare Isotope Beams (FRIB) in Michigan. FRIB is a nuclear science user facility that came online in the summer of 2022. FRIB’s users investigate nuclear properties across the nuclear landscape and require a variety of different experimental setups throughout the year. Since parameters such as beam composition and energy change between experiments in a given detector, the users of the facility need to train new ML models afresh for each experiment. Pretrained models can be quickly adjusted and adapted for users' desired use cases, reducing the burden on experimentalists. This work builds on industry-standard and industry-leading ML software and libraries, such as pytorch and tensorflow, and the pretrained models are openly available to the community, together with scripts for fine-tuning them for specific applications. The models are also integrated into standard software used by domain scientists. This work brings together experimental nuclear physicists, computer scientists and data scientists to create a tight-knit collaboration across disciplines. Undergraduate students play a central role in executing the research agenda. By engaging them in cutting-edge research in partnership with physicists at a national facility, students are prepared for impactful careers in STEM, both in academia and in the broader workforce.This award by the Office of Advanced Cyberinfrastructure is jointly supported by the Physics at the Information Frontier program in the Division of Physics within the Directorate for Mathematical and Physical Sciences.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.