Multi-target active subspaces generated using a neural network for computationally efficient turbulent combustion kinetic uncertainty quantification in the flamelet regime
Multi-target active subspaces generated using a neural network for computationally efficient turbulent combustion kinetic uncertainty quantification in the flamelet regime
复制标题
DOI：
10.1016/j.combustflame.2023.113015
复制
发表时间：
2023-12
影响因子：
4.4
通讯作者：
Benjamin C. Koenig;Sili Deng
中科院分区：
文献类型：
--
作者：
Benjamin C. Koenig;Sili Deng
通过燃烧模拟传播动力学模型中的不确定性可以提供有关模型可靠性和准确性的重要指标，但仍然是一个具有挑战性且数值昂贵的问题，特别是对于大型动力学模型和昂贵的湍流燃烧模拟。以前已经应用了各种替代模型和降维技术，以降低燃烧模拟中前向不确定性传播的成本，但这些通常仅限于具有标量解目标的低维、简单燃烧情况。在当前的工作中，我们开发了一个神经网络加速框架，用于识别低维主动动力学子空间，该子空间适用于火焰表的整个温度解空间，并且可以捕获动力学不确定性的混合分数和应变率相关效应。然后，我们通过在雷诺平均桑迪亚火焰 D 模拟中使用基于 217 个反应的甲烷燃烧化学模型的概念验证、基于火焰的应用程序，展示了该框架所实现的计算节省。通过利用主动子空间方法的大维压缩和低成本缩放，将初始降维梯度采样卸载到层流火焰模拟上，并使用专门设计的神经网络加速梯度采样过程，我们能够仅使用七个扰动解来估计湍流火焰解空间中的温度不确定性分布，精度高达 70−85%。此外，由于它完全发生在小火焰表内，因此识别缩减子空间的成本不会与湍流燃烧模型的成本成比例，这是该框架未来应用于更大规模和更复杂的湍流燃烧应用的一个有前途的特征。
Propagating uncertainties in kinetic models through combustion simulations can provide important metrics on the reliability and accuracy of a model, but remains a challenging and numerically expensive problem especially for large kinetic models and expensive turbulent combustion simulations. Various surrogate model and dimension reduction techniques have previously been applied in order to reduce the cost of forward uncertainty propagation in combustion simulations, but these are often limited to low-dimensional, simple combustion cases with scalar solution targets. In the current work, we developed a neural network-accelerated framework for identifying a low-dimensional active kinetic subspace that applies to the entire temperature solution space of a flamelet table and can capture the mixture fraction and strain rate dependent effects of the kinetic uncertainty. We then demonstrated the computational savings enabled by this framework through a proof-of-concept, flamelet-based application in a Reynolds-averaged Sandia Flame D simulation using a chemical model for methane combustion with 217 reactions. By leveraging the large dimensional compression and low-cost scaling of the active subspace method, offloading the initial dimension reduction gradient sampling onto the laminar flamelet simulations, and accelerating the gradient sampling process with a specifically designed neural network, we were able to estimate the temperature uncertainty profiles across the solution space of the turbulent flame with strong accuracy of 70− 85% using just seven perturbed solutions. Additionally, as it occurs entirely within the flamelet table, the cost of identifying the reduced subspace does not scale with the cost of the turbulent combustion model, which is a promising feature of this framework for future application to larger-scale and more complex turbulent combustion applications.