Scholarlink

<p>Modeling metal forming processes, and specifically ring rolling as considered in this research, is crucial for process design and the development of new product variants. Although finite element simulations and analytical modeling are common practices to perform process predictions, they can still present limitations such as high computational time, adherence to reality, and modeling accuracy. Trying to address all the above, this research proposes a novel generative machine learning solution for the creation of accurate and physics-compliant synthetic data for the ring rolling process, enabling almost real time and accurate process modeling within latent space. The proposed data generation method is based on multivariate generative adversarial network (GAN) combined with a physics-guided neural network (PGNN) to incorporate constraints relevant to the ring rolling process and integrate them into an auxiliary loss granting physical consistency. The proposed GAN architecture is then annotated as a physics-guided auxiliary GAN (PG-A-GAN). The physical constraints introduced are relevant to an analytical slip-line model for the force, a surrogate model for the torque, volume consistency, and an analytical model for the time–diameter relationship. Considering the duality of the physical modeling approach for the force and torque, where data-driven machine learning is stirred by analytical models, the proposed approach is defined as analytical transfer learning. Findings reveal that employing PGNNs in the GANs learning process improves physical loss in data generation, combined with a slight increase in data distribution similarity with respect to experimental instances. To enable modeling of specific and customized rolling instances, a condition based on the final sought ring geometry was introduced, within an auxiliary classifier generative adversarial network (ACGAN) framework. The proposed architecture allows the generation of multivariate, physically constrained, rolling time series and highlights the feasibility of such a modeling approach within the latent space and might be extended to other manufacturing processes by adapting modeling and retraining the proposed architecture.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Synthetic process modeling for ring rolling via physics-guided GANs and analytical transfer learning

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月21日

查看全文
  • Johannes Seitz,
  • Qinwen Wang,
  • Luca Quagliato,
  • Tobias Moser,
  • Taeyong Lee,
  • Alexander Brosius,
  • Bernd Kuhlenkötter

摘要

Modeling metal forming processes, and specifically ring rolling as considered in this research, is crucial for process design and the development of new product variants. Although finite element simulations and analytical modeling are common practices to perform process predictions, they can still present limitations such as high computational time, adherence to reality, and modeling accuracy. Trying to address all the above, this research proposes a novel generative machine learning solution for the creation of accurate and physics-compliant synthetic data for the ring rolling process, enabling almost real time and accurate process modeling within latent space. The proposed data generation method is based on multivariate generative adversarial network (GAN) combined with a physics-guided neural network (PGNN) to incorporate constraints relevant to the ring rolling process and integrate them into an auxiliary loss granting physical consistency. The proposed GAN architecture is then annotated as a physics-guided auxiliary GAN (PG-A-GAN). The physical constraints introduced are relevant to an analytical slip-line model for the force, a surrogate model for the torque, volume consistency, and an analytical model for the time–diameter relationship. Considering the duality of the physical modeling approach for the force and torque, where data-driven machine learning is stirred by analytical models, the proposed approach is defined as analytical transfer learning. Findings reveal that employing PGNNs in the GANs learning process improves physical loss in data generation, combined with a slight increase in data distribution similarity with respect to experimental instances. To enable modeling of specific and customized rolling instances, a condition based on the final sought ring geometry was introduced, within an auxiliary classifier generative adversarial network (ACGAN) framework. The proposed architecture allows the generation of multivariate, physically constrained, rolling time series and highlights the feasibility of such a modeling approach within the latent space and might be extended to other manufacturing processes by adapting modeling and retraining the proposed architecture.