Design of a Superheat Prediction Model for Continuous Casting Molten Steel Based on fi‐PINN

To control the superheat, a key parameter in the continuous casting process, and stabilize it at a predefined optimal value, it is necessary to predict the superheat corresponding to the casting start temperature. This article proposes a molten steel superheat prediction model comprising two submodels for ladle and tundish exit temperatures. The two submodels are developed using a hybrid approach combining both mechanistic and data‐driven methods, with a serial connection in structure and end‐to‐end training for data backpropagation. To improve model accuracy by capturing the time‐varying thermal properties of the working layer, this article proposes a forward‐inverse physics‐informed neural network (fi‐PINN) based on the traditional PINN algorithm. The fi‐PINN simultaneously solves partial differential equations and estimates parameters. Production data from multiple casting sequences at a steel plant are used to validate the model's performance. The results show that the proposed model improves the accuracy by ≈3.07 and 1.62 K, compared to the mechanistic and data models that do not account for the time‐varying thermal properties. Furthermore, the model's ability to capture the variation of the working layer's thermal properties provides strong physical interpretability.