Abstract:

Microreactors, as key devices for process intensification, have been applied in several fields. However, the lack of a microreactor module in most process simulation software has hindered its application in industry to some extent. In this paper, based on transfer learning, a product yield prediction model for Aspen microreactors applicable to large gas-liquid ratio is proposed, and the accuracy of the model is verified by taking the process of synthesizing dodecylbenzene sulfonic acid by large gas-liquid sulfonation in a microreactor as an example. First, 38 sets of yield data of dodecylbenzene sulfonic acid are collected in a T-type microreactor through sulfonation experiments with gas-liquid ratios ［(2000∶1)—(3000∶1)］, which serve as the target domain for transfer learning. A preliminary product yield prediction model for the process is developed based on the microchannel annular flow characteristics using the flat push-flow reactor (PFR) module of Aspen Plus, and 29700 sets of source domain data are generated. Considering the features such as fluid dynamics and microchannel structure, this study adopts and adapts the conditional adversarial domain adaptation network (CADAN) in transfer learning, including the adoption of a deep ReLU network architecture and optimization of the adversarial loss function. Subsequently, the feature extractor is trained using simulated data and the conditional adversarial domain adaptation is trained using experimental data. The final model fit coefficient (R²) can reach 0.9346, which is improved by 14.6% compared to the artificial neural network and 98.18% compared to the PFR model. Meanwhile, the model maintains an R² greater than 0.78 at 20% noise level, and the root mean square error is as low as 5.07, which is better than that of the artificial neural network under the same conditions, showing high prediction accuracy and strong robustness.

Key words: microreactor, process modeling, Aspen Plus, transfer learning, algorithm

摘要：

微反应器作为过程强化的关键设备，已在多个领域应用。然而，多数流程模拟软件缺乏微反应器模块，在一定程度上阻碍了其在工业中的应用。本文提出了一种基于迁移学习构建的适用于大气液比的Aspen微反应器的产物收率预测模型，以微反应器中气液磺化合成十二烷基苯磺酸过程为例，验证了该模型的准确性。首先，通过大气液比［（2000∶1）~（3000∶1）］磺化实验，在T形微反应器中收集38组十二烷基苯磺酸的收率数据，作为迁移学习的目标域。基于微通道环状流特征采用Aspen Plus的平推流反应器（PFR）模块初步建立该过程的产物收率预测模型，并生成29700组源域数据。考虑流体动力学和微通道结构等特征，本文采用迁移学习中的条件对抗域适应网络（CADAN）并对其进行调整，包括采用深度ReLU网络架构及优化对抗损失函数。随后利用模拟数据训练特征提取器，并利用实验数据进行条件对抗域适应训练。最终建立的模型拟合系数（R²）可达0.9346，相较人工神经网络提升了14.6%，较PFR模型提升了98.18%。同时，该模型在20%的噪声水平下仍保持R²大于0.78，均方根误差低至5.07，优于同等条件下的人工神经网络，显示出较高的预测精度和强鲁棒性。

关键词: 微反应器, 过程建模, Aspen Plus, 迁移学习, 算法