用于电化学流动反应器快速替代模型的物理信息神经网络
Physics-Informed Neural Networks as Fast Surrogate Models for Electrochemical Flow Reactors
浏览论文内容
中文总结 AI 辅助
本研究提出物理信息神经网络（PINN），将其作为电化学流动反应器的快速替代模型，经验证精度高、速度快，为低计算成本的数字孪生建模提供基础。
中文摘要 AI 辅助
本研究提出一种物理信息神经网络（PINN），用于建模包含扩散、迁移、对流及非线性阳极巴特勒-沃尔默动力学的瞬态二维电化学流动反应器。该模型通过将控制输运方程及所有初始条件、边界条件嵌入复合损失函数进行训练，无需标记浓度数据。输入为空间与时间坐标，以及阳极过电位、温度、入口浓度、最大流速和扩散率，使网络可预测宽操作域内的浓度场。与有限差分法求解结果的验证显示，典型瞬态及近稳态案例吻合度高，在整个条件域内的平均时空相对误差为$(9.99 \times 0.65)\times10^{-3}$（亚百分比水平）。PINN的推理速度比传统有限差分求解器快5.34倍，运行时间缩短81.3%。泛化测试进一步表明，该替代模型在域内边界聚焦采样及受控外推下仍保持鲁棒性，不过阳极过电位因对界面动力学的指数效应成为最具挑战性的参数。结果表明，物理信息神经网络可作为电化学输运问题精准且高效的参数化替代模型，为电化学流动反应器的低计算成本数字孪生建模奠定基础。
英文摘要
This work presents a physics-informed neural network (PINN) for modeling a transient two-dimensional electrochemical flow reactor with diffusion, migration, convection, and nonlinear anodic Butler--Volmer kinetics. The model is trained without labeled concentration data by embedding the governing transport equation and all initial and boundary conditions into a composite loss function. Spatial and temporal coordinates together with anodic overpotential, temperature, inlet concentration, maximum flow velocity, and diffusivity are used as inputs, allowing the network to predict concentration fields over a broad operating domain. Validation against finite-difference-based solutions shows strong agreement for representative transient and near-steady cases, with a mean relative space--time error of $(9.99 \pm 0.65)\times10^{-3}$ (sub-percent level) across the conditioning domain. PINN inference is faster than a traditional finite difference solver by a factor of $5.34$, thus reducing runtime by $81.3\%$. Generalization tests further show that the surrogate remains robust under in-domain boundary-focused sampling and controlled extrapolation, although anodic overpotential is the most challenging parameter due to its exponential effect on interfacial kinetics. The results indicate that physics-informed neural networks can serve as accurate and efficient parametric surrogates for electrochemical transport problems and provide a foundation for low-computational-cost digital-twin modeling of electrochemical flow reactors.
