Skip to Content

CrystalMind: A surrogate model for predicting 3D models with recrystallization in open-die hot forging including an optimization framework

This paper presents CrystalMind, a surrogate model designed for fast and accurate 3D prediction of deformation and recrystallisation of workpieces undergoing open die hot forging in 3D. Using machine learning, specifically MLP (Multi-Layer Perceptron) and PointNET++ architectures, CrystalMind significantly reduces computation time compared to traditional approaches such as FEM simulations.


Type of Work: Scientific Publication

Main Author: Jan Petrik

Affiliation: ETH Zurich

Co-Authors: Syed Irtiza Ali, Martin Feistle, Markus Bambach

Date: 1st December 2023

Journal: Mechanics of Materials

Online: ScienceDirect

CrystalMind uses a combination of FEM simulation for data generation and machine learning for recrystallisation and deformation prediction, focusing on the open-die hot forging process. Model inputs include a pre-stroke 3D part geometry and a forging vector detailing the forging strategy. Through data pre-processing and the use of MLP or alternatively PointNET++ based machine learning architectures, the model demonstrates the ability to accurately predict post-forging 3D shapes and recrystallisation approximately 240,000 times faster than traditional FEM simulations. In addition, the model's compliance with the volume conservation condition demonstrates adherence to the inherent law of physics.
Key findings from the use of CrystalMind demonstrate its ability to drastically reduce the computational resources and time required to simulate forging processes, with MLP proving to be 36 times faster than PointNET++, achieving an average computational time of 5ms per run, compared to around 20 minutes for FEM. The model ensures volume preservation and limits recrystallisation error to less than 2%. In addition, its integration with an optimisation algorithm based on customised dual annealing enables rapid optimisation of forging strategies, representing a significant advance over traditional methods by completing optimisation processes in minutes instead of years.


  • Image 1
    The representation of both the input and output from the Abaqus simulation is showcased, highlighting the mesh deformation and recrystallization process. This information was instrumental in training the AI-based model known as CrystalMind. It's also important to mention that due to axial symmetries, the simulation only required one quarter of the workpiece.
  • Image 1
    Training workflow of CrystalMind. The architecture accepts as input a 3D model and a forging strategy defined as a 4D vector. Afterwards, it reconstructs the 3D modelbased on this input, incorporating a forging stroke according to the 4D vector. The resulting prediction is compared with a ground truth 3D model using two loss functions basedon mean absolute error (MAE). Through multiple epochs, the total loss is backpropagated to the architecture for training.
  • Video of a GUI demonstration built on CrystalMind, where users can specify the forging strategy and input geometry. In just a few milliseconds, they receive feedback on the recrystallization and deformation outcomes of a workpiece.