Skip to Content

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

This study presents a method for microstructure control in metal forging that combines Model Predictive Control (MPC) with a newly developed machine learning model called DeepForge. Using 1D convolutional neural networks and gated recurrent units, DeepForge predicts microstructural changes from surface temperature measurements, providing a novel approach to improving forging quality and precision. By integrating real-time control via MPC, this research demonstrates the ability to dynamically adjust process parameters for optimised microstructural results.


Type of Work: Scientific Publication

Main Author: Jan Petrik

Affiliation: ETH Zurich

Co-Authors: Markus Bambach

Date: 25th February 2024

Journal: Journal of Manufacturing Processes

Online: Arxiv

The research employs a sophisticated methodology, combining finite element (FE) simulation with machine learning to create DeepForge. This hybrid model uses surface temperature data to predict microstructural outcomes such as grain size and recrystallisation across a workpiece. The AI-driven architecture consists of 1D convolutional neural networks and gated recurrent units to process temporal and spatial data. In addition, the study utilises a comprehensive dataset generated from simulated three-stroke forging processes, providing a robust basis for DeepForge's predictions.
Key results demonstrate DeepForge's ability to predict microstructural changes with good accuracy, achieving a mean absolute error of only 0.4%. Furthermore, when integrated with MPC, the system successfully adjusts forging parameters in response to temperature disturbances, maintaining target grain sizes. By enabling dynamic process control, DeepForge paves the way for improvements in material properties and part quality. Finally, the results of dynamic process control were experimentally verified, demonstrating the applicability of DeepForge in a real-world scenario.


  • Image 1
    Graphical abstract showcasing DeepForge's capability to predict the microstructure of a workpiece during forging, using only surface temperature measurements. Additionally, it highlights the integration of Model Predictive Control, enabling real-time adjustment of process parameters to attain specific microstructural characteristics.
  • Image 1
    DeepForge's training workflow uses past surface temperatures and a forging strategy as inputs, defaulting initial temperatures to 0 if unavailable. It reconstructs 2D microstructure arrays of the part, considering the forging effects. The model is refined through comparison with actual data using Mean Absolute Error (MAE).
  • Image 1
    Qualitative results that show a prediction of the workpiece microstructure based exclusively on surface temperatures. The forecasted microstructure is displayed on the left, with the percentage error compared to the actual ground truth data presented on the right.