III: Medium: Collaborative Research: Deep Learning for In Situ Analysis and Visualization

随着科学家们预计亿级计算的好处，缺乏新的解决方案来处理大规模数据和校准模拟参数已成为进一步加快科学发现的重大障碍。本项目的目标是开发一种新的基于深度神经网络的端到端数据分析和特征提取工作流，以帮助计算科学家解决三个主要挑战：(1)识别重要的模拟参数并生成用于分析的基本数据；(2)将模拟数据转换为紧凑的特征表示以传达最大的洞察力；(3)设计可扩展的可视化算法，并结合大规模模拟来收集对其科学问题的见解。该团队将与喷气发动机设计、气候模型、心脑血管流动、超导和聚变能源领域的科学家合作，演示深度学习技术如何帮助从大量模拟数据中提取特征，并在巨大的模拟参数空间中导航。通过暑期实习和项目合作，这项研究将为研究生和本科生创造机会，包括来自代表性不足群体的学生，与顶尖科学家一起参与关键研究计划。通过计划举办的“可视化的深度学习”年度暑期班，研究成果将使可视化研究人员和更广泛的社区能够将开发的深度学习技术的原则和实践纳入其中。研究小组将开发一个全面的分析框架，其中包括一套最先进的深度学习技术，用于现场处理和分析大规模科学模拟数据。该框架将由三个紧密结合的部分组成：(1)仿真参数分析和数据约简，(2)数据和特征的后处理分析，(3)现场工作流优化。对于第一部分，将开发方法来辅助模拟代理创建、参数空间探索和总体模拟的比较分析。对于第二个组成部分，将开发深度学习技术，以便从数据中学习特征，以便对代表进行互动探索，并在空间和时间领域扩大减少的模拟产出。对于第三个组成部分，将为特征检测、工作量估计和特征计算代理制定现场解决方案。该框架将使用四种类型的量化度量进行评估：数据缩减率、数据、特征和图像级别的误差度量、可伸缩性度量以及与训练和测试数据的交叉验证。该团队将在喷气发动机、气候、心脑血管流动、超导和聚变能源等领域与科学家密切合作。领域科学家将发挥关键作用，使研究团队能够了解他们的应用程序的要求，并评估这项研究的结果。该项目的传播计划将面向更广泛的受众，包括学生、从业者和领域科学家，以增强他们对深度学习对可视化的价值的理解和欣赏。该团队将发布这项研究产生的开源软件、预先训练的模型以及训练和测试数据，包括用于特征学习的自动编码器、DNN辅助的参数空间探索、基于CNN的特征提取和跟踪以及基于负载平衡的深度预测模型。该奖项包括来自计算机与硬件部门信息与智能系统、软件与硬件系统部门的信息集成与信息学计划的资金。这一奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

As scientists anticipate the benefits of exascale computing, the lack of novel solutions to process data at scale and calibrate the simulation parameters has become a significant roadblock to further accelerating scientific discovery. The goal of this project is to develop a new end-to-end data analysis and feature extraction workflow based on deep neural networks to help computational scientists address three major challenges: (1) identify important simulation parameters and generate the essential data for analysis, (2) transform the simulation data to compact feature representations to convey the most insight, and (3) design scalable visualization algorithms coupled with large-scale simulations to glean insight into their scientific problems. Working with domain scientists in jet engine design, climate models, cardio/cerebrovascular flow, superconductivity, and fusion energy, the team will demonstrate how deep learning techniques can help extract features from vast amounts of simulation data and navigate in the huge simulation parameter space. Through summer internships and project collaborations, this research will create opportunities for graduate and undergraduate students, including students from underrepresented groups, to participate in key research initiatives with leading scientists. Through the planned annual summer school on "Deep Learning for Visualization," the research results will enable visualization researchers and a broader community to incorporate the principles and practice of deep learning techniques developed. The research team will develop a comprehensive analysis framework that encompasses a suite of state-of-the-art deep learning techniques for in situ processing and analysis of large-scale scientific simulation data. The framework will consist of three tightly-integrated components: (1) analysis of simulation parameters and data reduction, (2) post-analysis of data and features, and (3) in situ workflow optimization. For the first component, methods will be developed to assist simulation surrogate creation, parameter space exploration, and comparative analytics of ensemble simulations. For the second component, deep learning techniques will be developed to learn features from data for interactive exploration of representatives and to upscale reduced simulation output in the spatial and temporal domains. For the third component, in situ solutions will be developed for feature detection, workload estimation, and feature computation surrogates. The framework will be evaluated using four types of quantitative metrics: data reduction ratio, data-, feature-, and image-level error measures, scalability measures, and cross-validation with training and testing data. The team will work closely with scientists in the domains of jet engine, climate, cardio/cerebrovascular flow, superconductivity, and fusion energy. The domain scientists will play a critical role in enabling the research team to understand the requirements of their applications and to evaluate the outcomes of this research. The project's dissemination plan will address a much broader audience, including students, practitioners, and domain scientists, to enhance their understanding and appreciation of the value of deep learning for visualization. The team will release open-source software, pre-trained models, and training and test data generated from this research, including auto-encoder for feature learning, DNN-assisted parameter space exploration, CNN-based feature extraction and tracking, and load-balancing based deep predictive models.This award includes funding from the Information Integration & Informatics Program in the Division of Information & Intelligent Systems, Software & Hardware Systems Program in the Division of Computer & Computing Foundations, and the NSF Office of Advanced Cyberinfrastructure.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.