物理信息神经网络代理求解燃烧反应动力学

A Physics-informed Neural Network for Solving Combustion Reaction Kinetics

现阶段基于神经网络构建化学反应动力学的代理模型被认为是加速燃烧数值计算的重要方法。纯数据驱动神经网络长期存在过于依赖数据采样、鲁棒性和泛化性难以保证等问题。为此,本文在训练化学反应动力学代理模型时引入了质量作用定律、质量守恒、能量守恒、元素守恒等物理信息的约束。在零维自燃和二维本生灯火焰的数值计算中,物理信息神经网络在保证2.0~4.7倍加速比的同时,抑制了神经网络代理模型的预测偏差,并且根据输运方程源项误差的理论估计给出了降低代理模型误差、提高模型鲁棒性的相关建议。

Recently,surrogate models for solving chemical reaction kinetics based on neural net-works have been considered to be critical for accelerating simulations of combustion reaction flows.Data-driven neural networks have long been plagued by issues such as an over-reliance on data sam-pling and dificulty in ensuring robustness and generalization.To address these issues,the training of neural networks is constrained by physical principles in the present study,including the law of mass action,and the conservation of mass,energy,and elements.Compared with data-driven neural network surrogate models,the physics-informed neural network accelerates computations by 2.0~4.7 times while suppressing prediction errors in the numerical simulations of o-D autoignitions and a 2-D Bunsen flame.Finally,based on theoretical estimates of the source term error in combustion reaction flow simulations,this study introduces several suggestions to reduce prediction errors and improve model robustness.