Predictive control for a single-blow cold upsetting using surrogate modeling for a digital twin

Abstract

In the realm of forging processes, the challenge of real-time process control amid inherent variabilities is prominent. To tackle this challenge, this article introduces a Proper Orthogonal Decomposition (POD)-based surrogate model for a one-blow cold upsetting process in copper billets. This model effectively addresses the issue by accurately forecasting energy setpoints, billet geometry changes, and deformation fields following a single forging operation. It utilizes Bézier curves to parametrically capture billet geometries and employs POD for concise deformation field representation. With a substantial database of 36,000 entries from 60 predictive numerical simulations using FORGE® software, the surrogate model is trained using a multilayer perceptron artificial neural network (MLP ANN) featuring 300 neurons across 3 hidden layers using the Keras API within the TensorFlow framework in Python. Model validation against experimental and numerical data underscores its precision in predicting energy setpoints, geometry changes, and deformation fields. This advancement holds the potential for enhancing real-time process control and optimization, facilitating the development of a digital twin for the process.