Toluene diisocyanate (TDI) is an important raw material to produce a flexible polyurethane foam, and the demand for TDI is growing as the polyurethane market is driven by high demand. The cold phosgenation reactor plays a vital role in the production of TDI, in that the overall reaction selectivity is determined and the most of by-product urea, which is critical to the entire downstream process, is produced inside the reactor. Therefore, the optimal design of the cold phosgenation reactor is very important to improve the overall process efficiency and operability of the TDI production process. In this research, we develop a framework for designing and optimizing TDI reactors through Bayesian optimization methods with design parameters including the diameter of two inlet nozzles, the angle between the nozzles, the size of the mixing zone, and the ratio of the converging-diverging nozzle. A comprehensive 3-dimensional computational fluid dynamics (CFD) reactor model is incorporated into the Gaussian process (Kriging) to construct a surrogate model, whose posterior is subsequently updated with new sample points searched by the acquisition function evaluated within the Bayesian optimization algorithm. As a result, the optimal design is obtained, and the urea selectivity is reduced by 11.6% compared with the basic design scheme. Compared to the 6 × 10⁷ simulations required for full grid search, only 61 function evaluations were performed to attain the optimum, demonstrating that the proposed framework will help efficiently achieve the optimal design of the expensive CFD reactor models that demand a high computational cost and time for evaluation.

甲苯二异氰酸酯（TDI）是生产软质聚氨酯泡沫的重要原料，随着聚氨酯市场的需求旺盛，对TDI的需求也在增长。冷光气化反应器在TDI的生产中起着至关重要的作用，因为确定了总的反应选择性，并且在反应器内部产生了对整个下游过程至关重要的大部分副产物尿素。因此，冷光气反应器的优化设计对于提高TDI生产工艺的整体工艺效率和可操作性非常重要。在这项研究中，我们开发了一个通过贝叶斯优化方法设计和优化TDI反应器的框架，其设计参数包括两个入口喷嘴的直径，喷嘴之间的角度，混合区的大小，和收缩喷嘴的比例。将一个综合的3D计算流体动力学（CFD）反应堆模型合并到高斯过程（Kriging）中，以构建一个替代模型，该模型的后验随后用在贝叶斯优化算法中评估的采集函数搜索的新样本点进行更新。结果，获得了最佳设计，并且与基本设计方案相比，尿素选择性降低了11.6％。与6相比 其后验随后用在贝叶斯优化算法中评估的采集函数搜索的新样本点进行更新。结果，获得了最佳设计，并且与基本设计方案相比，尿素选择性降低了11.6％。与6相比 其后验随后用在贝叶斯优化算法中评估的采集函数搜索的新样本点进行更新。结果，获得了最佳设计，并且与基本设计方案相比，尿素选择性降低了11.6％。与6相比 ×全网格搜索所需的 10 ^7个仿真，仅执行了61个功能评估以达到最佳，表明所提出的框架将有助于有效地实现昂贵的CFD反应堆模型的优化设计，而该模型需要较高的计算成本和评估时间。