Roll pass design for round and square sections using an informed artificial neural network

Abstract. Nowadays, the roll pass design process for hot rolling mills of full sections like round and square sections is carried out by FEM simulations or analytical calculations. Both techniques require a highly iterative method with numerous iterations required to determine a satisfying pass design solution. In the present work, a neural network is trained for the process of roll pass design for round and square sections. The steps of data generation and training are combined by the formulation of a custom loss function for the underlying physical problem. In this way, the neural network is informed about the metal forming problem, enabling the generation of the necessary training data to be carried out implicitly during training. This technique increases the flexibility and extensibility of the current approach, eliminating the necessity of external training data generation. The metal forming equations for the pass design problem are presented for the pass sequences round-oval-round, as well as square-diamond-square, which are included in custom loss functions for the neural networks. The prediction results are compared to analytical iterative calculations, indicating very good agreements. Therefore, multiple iterations and computation time can be saved when the new approach of pass design is applied.