Adaptive Physics-informed Machine Learning Strategies for Turbulent Combustion Modeling

在应对气候变化和实现国家安全目标的任务中，燃烧装置的设计和优化至关重要。模拟可以通过在各种设计配置下对燃烧器进行快速虚拟测试，从而可以选择有前途的设计进行物理原型设计，从而在这项任务中发挥作用。然而，湍流燃烧模拟涉及求解大量的分子物种，这些分子物种是燃烧过程中产生和消耗的一部分。因此，燃烧模拟需要在超级计算机或计算集群上使用许多计算机处理器几个小时。这限制了有前途的计算机模型对实际设计和优化工作的有用性。这项工作有助于开发减少燃烧模型，减少模拟时间和所需的计算资源，同时保持准确性。为了应对湍流燃烧的计算成本过高，引入了基于物理的降阶模型。这些模型通常在“离线”阶段求解化学，将溶液存储在表中，然后插入表中的条目以检索“在线”阶段的化学状态。然而，这些查找表的使用受到维度的困扰，其中表的大小和插值复杂性随着控制变量的数量呈指数增长。因此，这些查找表仅限于使用少数控制变量的情况，从而阻碍了它们在许多实际燃烧装置中的应用。这项工作旨在通过开发机器学习策略来解决这个问题，以便在物理推导的低维流形中有效地学习燃烧物理。这将通过引入自适应的机器学习模型来实现，这些模型与基本的物理定律一致，并适用于高维燃烧状态空间。这项工作将使模拟达到目前使用传统制表技术无法实现的逼真程度，因此，将有助于设计和开发清洁高效的燃烧技术。此外，本研究中开发的工具对于科学机器学习的广泛领域至关重要，机器学习将继续对许多物理系统的建模产生越来越重要的影响。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。

The design and optimization of combustion devices is crucial in the mission to combat climate change and achieve national security goals. Simulations can play a role in this mission by enabling rapid virtual testing of combustors at various design configurations, so that promising designs can be selected for physical prototyping. However, turbulent combustion simulations involve solving for a large number of molecular species that are produced and consumed as part of the combustion process. Due to this, combustion simulations require the use of many computer processors for several hours on supercomputers or compute clusters. This limits the usefulness of promising computer models for practical design and optimization endeavors. This work contributes to the ongoing quest to develop reduced combustion models that decrease the simulation times and required computing resources, yet preserve accuracy.In response to the excessive computational costs of turbulent combustion, physics-based reduced-order models have been introduced. These models often solve chemistry in an “offline” phase, store the solution in a table, and then interpolate the table’s entries to retrieve the chemical state during the “online” phase. The use of these lookup tables, however, suffers from the curse of dimensionality, wherein the size of the table and the interpolation complexities increase exponentially with the number of control variables. As a result, these lookup tables are limited to situations that employ a few control variables, thus preventing their application to many practical combustion devices. This work aims to address this problem by developing machine learning strategies to efficiently learn combustion physics within physically derived low-dimensional manifolds. This will be achieved by introducing machine learning models that are adaptive, consistent with the underlying physical laws, and suitable for high-dimensional combustion state spaces. This work will enable simulations at levels of fidelity that are currently impossible to perform using traditional tabulation techniques, and therefore, will aid in the design and development of clean and efficient combustion technologies. Furthermore, the tools developed in this study will be vital to the broad area of scientific machine learning, which continues to have an increasingly important impact on the modeling of many physical systems.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.