Reactors in metallurgy are often complex non-linear physicochemical reaction systems. The reaction processes inside these reactors directly affect the quality and performance of the reaction products. Because the reaction process is complex and often takes place at high temperature and pressure, it is difficult to obtain three-dimensional reaction field information, including flow field, gas-phase reactant concentration, particle concentration distribution and other data, which also makes it difficult to achieve intelligent control and optimal regulation of the reaction process. To solve this problem, accurate information about the reaction field inside the reactor is needed to ensure that the flow and reaction process inside the reactor are within the controllable range. However, due to the often complex operating conditions inside the reactor, such as confinement, high temperature and pressure, and low visibility, the parameters and area range that can be measured by traditional measurement methods are limited, and are influenced by human factors, making it difficult to obtain accurate three-dimensional reaction field information. Therefore, obtaining multi-field coupled data inside the reactor remains a key scientific problem and a common problem for process optimization in general. In recent years, various data-driven methods, such as machine learning and deep learning, have been widely used for data prediction of industrial processes. However, traditional deep learning methods usually rely on the input of historical operational data sets and ignore the reaction mechanism model, which has inherent deficiencies in terms of predictable data types and accuracy. At the same time, with the increasing sophistication of numerical simulation, it is possible to couple actual reaction mechanisms to accurately calculate the flow, heat transfer, and chemical reaction processes within a reactor. However, using only numerical simulation methods, the operation steps are cumbersome and the simulation speed is slow, which makes it difficult to achieve continuous, real-time and efficient prediction of the reactor interior. In order to solve the problems of "black box" and parameter optimization and control inside the reactor, we propose a method to visualize and control the three-dimensional reaction field dynamically and in real time in the rotary kiln and gas-based shaft furnace direct reduction process, which can significantly reduce the calculation time while ensuring the accuracy.