DeepForge: Leveraging AI for Microstructural Control in Metal Forming via Model Predictive Control

TL;DR

The paper tackles microstructure control in hot closed-die forging under disturbances by developing DeepForge, a soft-sensor that predicts full-field microstructure from measurable surface temperatures using a hybrid 1D CNN–GRU architecture. It integrates DeepForge with Model Predictive Control to adaptively adjust inter-stroke wait times and maintain a target grain size of $< 35\ \mu\mathrm{m}$ in a specified region, achieving a mean absolute error of $0.4\pm0.3\%$ for microstructure predictions. The approach is validated through FE simulations and experimental verification, demonstrating real-time capability (approximately $4\ \mathrm{ms}$ per stroke) and improved robustness to disturbances compared to traditional FE workflows. This work advances practical microstructure control in forging by enabling in-situ, data-driven adjustments based on surface-temperature measurements, with clear avenues to extend to more geometries and 3D modeling.

Abstract

This study presents a novel method for microstructure control in closed die hot forging that combines Model Predictive Control (MPC) with a developed machine learning model called DeepForge. DeepForge uses an architecture that combines 1D convolutional neural networks and gated recurrent units. It uses surface temperature measurements of a workpiece as input to predict microstructure changes during forging. The paper also details DeepForge's architecture and the finite element simulation model used to generate the data set, using a three-stroke forging process. The results demonstrate DeepForge's ability to predict microstructure with a mean absolute error of 0.4$\pm$0.3%. In addition, the study explores the use of MPC to adjust inter-stroke wait times, effectively counteracting temperature disturbances to achieve a target grain size of less than 35 microns within a specific 2D region of the workpiece. These results are then verified experimentally, demonstrating a significant step towards improved control and quality in forging processes where temperature can be used as an additional degree of freedom in the process.

Figures (10)

• Figure 1: A diagram illustrating the simulation that involves a three-stage hot closed-die forging process, in which a specimen is moved from a preheated oven to a designated temperature in transport time and then subjected to three forging strokes. Each forging stroke's time is indicated by the upsetting time, while the interval between two successive strokes is expressed by a parameter called wait time. Finally, the last wait time is between the final stroke and the quenching. It is also worth noting that the green arrows visualise the moments when the simulation outputs the microstructure of the workpiece.
• Figure 2: Visualization of the three-stroke hot forging FE-based simulation, where the colour bar corresponds to displacement in y denoted as $U_{y}$. The specimen consists of 20$\times$10 finite elements giving a total of 231 nodal values. It should also be noted that, due to axial symmetry, only one-half of the part is simulated.
• Figure 3: Example of an input-output pair, where the applied stroke is characterized with the wait time of 7.6 s and upsetting time of 0.13 s.
• Figure 4: Visualisation of the input required for DeepForge, where only the normalized surface temperature that can be measured in the real process is used to predict the full microstructure of a workpiece consisting of six 2D arrays.
• Figure 5: The general workflow of DeepForge in the train scenario involves taking surface temperatures, i.e. surface temperatures, from time moments $t-3$, $t-2$, $t-1$ along with a forging strategy as input. It is important to highlight that in instances where prior measurements are absent, such as when a part is freshly removed from an oven and undergoes deformation during the first forging stroke, the first two surface temperatures are assigned a default value of 0. This input information is then passed through DeepForge, which ultimately reconstructs 2D arrays expressing the microstructure of the part, incorporating the effects of the forging strategy applied. Finally, the reconstructed arrays are compared to the ground truth using the Mean Absolute Error (MAE) loss function. This loss is backpropagated and the entire network is trained over a number of epochs to optimise the performance of the model.