EAGER:Predictive Surrogate Modeling and Analysis of Radiative Heat transfer in Porous Media

多孔介质中的辐射传热是复杂而模糊的；然而，测量目标多孔介质的辐射特性是至关重要的。辐射特性的测量和预测对于涉及多孔材料结构的能源技术的模拟和设计至关重要，包括核反应堆中的球床、选择性激光烧结技术、太阳能吸收器、太阳能热化学反应器、生物组织、喷气发动机和航天飞行器的热障、催化燃烧的陶瓷泡沫等等。目前，预测随机充填地层的辐射特性需要大量耗时的射线追踪模拟。该项目用高效的基于机器学习的方法取代了这些计算，从而彻底改变了广泛的相关应用和底层技术。这一变革项目表明，代理模型可以可靠有效地近似和预测随机填充结构的辐射特性的概率分布函数。大型耗时的光线追踪蒙特卡罗模拟被基于机器学习方法的预测模型所取代。模型的输入是有关空隙、固体、边界条件和尺寸以及介质形状的物理结构的各种变量的统计数据。研究了用于数据拟合的各种学习模型，并对每种模型的准确性和计算成本进行了分析。数据采样、模型选择和模型拟合都是为了呈现准确、高效、可扩展和可推广的代理模型而设计的。研究了每个代理模型的采样、实验设计和模型拟合，以减少计算量，同时最大限度地减少数据收集和学习的成本。通过与直接蒙特卡罗模拟和先前建立的实验室实验的比较，验证了所提出模型的实际准确性。所提出的预测模型已应用于计算机断层扫描中，用于各种应用中多孔介质结构的推断。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。

Radiative heat transfer in porous media is complex and ambiguous; yet, measuring the radiative properties of a target porous media is of vital importance. Measurement and prediction of the radiative properties is critical for simulation and design of energy technologies that involve porous material structures, including pebble beds in nuclear reactors, selective laser sintering technology, solar absorbers, solar thermochemical reactors, biological tissues, thermal barriers for jet engines and space vehicles, ceramic foams for catalytic combustion and many more. At present, predicting radiative properties of randomly packed beds requires large time-consuming ray-tracing simulations. This project replaces these computations with efficient machine learning based methods to revolutionize a wide range of related applications and underlying technologies. This transformative project demonstrates that surrogate models can approximate and predict the probability distribution functions of radiative properties of randomly packed structures reliably and efficiently. Large time-consuming ray-tracing Monte Carlo simulations are replaced by predictive models based on machine learning methods. The inputs to the models are statistics of a wide range of variables pertaining to the physical configurations of void, solid, boundary conditions and dimensions and the medium shape. Various learning models are studied for data fitting, and an analysis of accuracy versus the cost of computation is performed for each. Data sampling, model selection and model fitting are all engineered to render surrogate models that are accurate, efficient, scalable and generalizable. Sampling, design of experiment and model fitting is studied for each surrogate model to reduce the computational load while minimizing the cost of data collection and learning. The practical accuracy of the proposed models is validated based on comparison with direct Monte Carlo simulations and previously established laboratory-based experiments. The proposed predictive models are applied in computed tomography for inference of porous media structures in various applications.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.