Abstract Dual-fuel (DF) compression ignition engines, which employ a high-reactivity pilot fuel (e.g. diesel or DME) to ignite a low-reactivity lean premixed charge (e.g. methane/air), have been proposed to meet stringent pollutant regulations. Due to the complex multiscale interaction among flow, chemistry and flames, DF combustion exhibits a complicated, multi-modal combustion regimes and is hence challenging to model. Ignition delay time (IDT), as one of the most important parameters, is typically considered to develop an understanding and modeling strategy for complex ignition processes. However, accurate calculations and measurements of the IDTs over a wide range of fuel blends, pressures and flow conditions is a time-consuming, complicated procedure. While several physics-based IDT models have been proposed for single fuel ignition, they are subject to some limitations in DF scenarios. In this work, two different supervised Machine Learning methods: a glass box – High Dimensional Model Representation (HDMR) and a black box – Convolutional Neutral Network (CNN) are employed to seek an accurate and efficient prediction of the IDTs of DF. First, the underlying mechanisms of DF interaction during the ignition process are investigated. The results show that the DF ignition process is highly complex, involving negative-temperature coefficient (NTC) behavior, two-stage ignition, and multiple combustion modes and transition. Then, data needed to train HDMR and CNN is generated by a large number of transient counterflow and homogeneous reactor calculations covering DF engine conditions. The trained HDMR and CNN models are tested with numerical and experimental databases. The results show that both HDMR and CNN can capture the features of DF ignition and correctly predict IDTs, even for predictions outside the ranges of parameters used for learning. Compared to CNN, HDMR is more favorable due to its relatively weak dependence on the size of training data and its ability to assess the sensitivity of IDTs to input variables. The sensitivity analysis suggests that the mixing rate between the pilot fuel and the main fuel plays a critical role in affecting the DF ignition. The HDMR and/or CNN models are seen as promising alternatives to time-consuming experimental measurements or numerical calculations of IDTs.

中文翻译：

---

双燃料发动机条件下点火延迟时间的机器学习

摘要 双燃料 (DF) 压燃式发动机采用高反应性引燃燃料（例如柴油或 DME）来点燃低反应性贫预混料（例如甲烷/空气），已被提议满足严格的污染物法规。由于流动、化学和火焰之间复杂的多尺度相互作用，DF 燃烧表现出复杂的多模式燃烧状态，因此建模具有挑战性。点火延迟时间 (IDT) 作为最重要的参数之一，通常被认为是开发复杂点火过程的理解和建模策略。然而，在各种混合燃料、压力和流量条件下准确计算和测量 IDT 是一个耗时且复杂的过程。虽然已经提出了几种基于物理的 IDT 模型用于单燃料点火，它们在 DF 场景中受到一些限制。在这项工作中，采用了两种不同的监督机器学习方法：玻璃盒——高维模型表示（HDMR）和黑盒——卷积中性网络（CNN）来寻求对 DF 的 IDT 的准确有效预测。首先，研究了点火过程中DF相互作用的潜在机制。结果表明，DF点火过程非常复杂，涉及负温度系数（NTC）行为、两阶段点火以及多种燃烧模式和转变。然后，训练 HDMR 和 CNN 所需的数据是通过覆盖 DF 发动机条件的大量瞬态逆流和均质反应堆计算生成的。经过训练的 HDMR 和 CNN 模型使用数值和实验数据库进行测试。结果表明，HDMR 和 CNN 都可以捕获 DF 点火的特征并正确预测 IDT，即使是在用于学习的参数范围之外的预测。与 CNN 相比，HDMR 更有利，因为它对训练数据大小的依赖性相对较弱，并且能够评估 IDT 对输入变量的敏感性。敏感性分析表明，引燃燃料和主燃料之间的混合率在影响 DF 点火中起着关键作用。HDMR 和/或 CNN 模型被视为 IDT 耗时的实验测量或数值计算的有前途的替代方案。HDMR 更有利，因为它对训练数据大小的依赖性相对较弱，并且能够评估 IDT 对输入变量的敏感性。敏感性分析表明，引燃燃料和主燃料之间的混合率在影响 DF 点火中起着关键作用。HDMR 和/或 CNN 模型被视为 IDT 耗时的实验测量或数值计算的有前途的替代方案。HDMR 更有利，因为它对训练数据大小的依赖性相对较弱，并且能够评估 IDT 对输入变量的敏感性。敏感性分析表明，引燃燃料和主燃料之间的混合率在影响 DF 点火中起着关键作用。HDMR 和/或 CNN 模型被视为 IDT 耗时的实验测量或数值计算的有前途的替代方案。