BEGIN:VCALENDAR
VERSION:2.0
METHOD:PUBLISH
BEGIN:VEVENT
ORGANIZER;CN=ESTAD 2023:mailto:info@metec-estad.com
LOCATION:Room 28
SUMMARY:A neural network for rapid roll pass design of full sections
DESCRIPTION:Classical approaches of roll pass design (RPD) for finished full sections require a two-step method. First, main grooves are designed according to a predefined elongation distribution to meet the requirements of the intended process in terms of temperature evolution, power demand and section tolerances. In the second step, the intermediate grooves are designed in order to fulfill prescribed filling ratios, taking into account the specific spread behaviour of the rolling process. This process is a trial-and-error method and often requires a high number of iterations to yield a satisfying solution for all passes. Extra complexity exists for diamond-diamond pass sequences in which a final square section is produced in few passes without predefined main grooves with similarities to the three-roll process, suffering from the same restrictions.
To speed up the design process and to initiate a framework for a fully-automated solution, the RPD problem is solved by a machine-learning method in the present study. Using the industrially approved RPD software MPC developed at our research group, we generated training data for the pass sequences round-oval-round, square-diamond-square and diamond-diamond-square in the two-roll process, as well as round-round and round-hexagon in the three-roll process, each for a range of roll diameters and total reductions.
The fully-connected neural network trained with these data is able to predict pass designs within the trained data ranges. Verification data was generated which does not coincide with the training data to test the network against overfitting. 
Model results are presented and cross-validated against the analytical RPD model.
The resulting ML pass design model is implemented in Python using the PyTorch library. Source code for including the parameters of the neural network will be published for third-party evaluation.
CLASS:PUBLIC
DTSTART:20230614T094000
DTEND:20230614T100000
END:VEVENT
END:VCALENDAR