Fast-acting artificial intelligence models for real-time digital twins of metal forming processes
DOI:
https://doi.org/10.37142/2076-2151/2025-1(54)141Keywords:
digital twin; finite element method; real time AI models; low latency inference; hybrid Conv3D–LSTM neural architectures; reinforcement learning; adaptive control; AI explainability; metal forming; energy optimization.Abstract
Kovalevskyy S. Fast-acting artificial intelligence models for real-time digital twins of metal forming processes
This paper presents a comprehensive methodology for constructing a digital twin of metal forming processes by combining classical finite element modeling with high‑speed artificial intelligence models. The study aimed to achieve inference latency of no more than one hundred milliseconds, accelerate the complete simulation cycle by more than fifty times compared to traditional FEM computations (initially anticipated as a threefold speed‑up), and attain predictive accuracy above ninety percent. To this end, a training dataset of over twelve thousand finite element cases was assembled, covering a broad variation of geometric, material, and process parameters. Based on these data, a hybrid neural‑network architecture was developed, alternating three‑dimensional convolutional layers and recurrent long short‑term memory blocks to reproduce complex stress and displacement fields using voxel tensors. Adaptive control was implemented via a reinforcement‑learning agent capable of dynamically adjusting press parameters during operation according to a multi‑component reward function that balances surface quality, energy consumption, and productivity. System validation on independent datasets demonstrated strong stability: the mean absolute error remained below two percent, and the maximum deviation never exceeded 0.08 units. Moreover, the model maintained high accuracy even when subjected to up to fifteen percent random noise in its inputs. Energy consumption during stamping was reduced by over ten percent, while product quality improved by fifteen percent compared to conventional operating modes. The resulting platform features scalable, modular architecture with well‑defined interfaces to manufacturing execution (MES) and enterprise resource planning (ERP) systems, and it provides transparency of AI modules through visualization of key input‑feature contributions. Future work will prioritize knowledge‑distillation techniques for deployment on embedded controllers and further optimization of computational resources.
References
Sahoo S., Lo C.-Y. Smart manufacturing powered by recent technological advancements. Journal of Manufacturing Systems. 2022. Vol. 64. P. 236–250. ISSN 0278-6125. DOI: https://doi.org/10.1016/j.jmsy.2022.06.008.
Petru C.-D., Crenganiș M., Breaz R.-E., Racz S.-G., Gîrjob C.-E., Drașovean P. Development of a digital twin for the ABB IRB 1200 robot in sheet metal forming processes. Procedia Computer Science. 2024. Vol. 242. P. 228–235. ISSN 1877-0509. DOI: https://doi.org/10.1016/j.procs.2024.08.253.
Ge Q., Li J., Lacasse S., Sun H., Liu Z. Data-augmented landslide displacement prediction using generative adversarial network. Journal of Rock Mechanics and Geotechnical Engineering. 2024. Vol. 16, № 10. P. 4017–4033. ISSN 1674-7755. DOI: https://doi.org/10.1016/j.jrmge.2024.01.003.
Li J., Tong Ch., Zhang R., Shi Zh., Lin J. A data-informed review of scientific and technological developments and future trends in hot stamping. International Journal of Lightweight Materials and Manufacture. 2024. Vol. 7, No. 2. P. 327–343. ISSN 2588-8404. DOI: https://doi.org/10.1016/j.ijlmm.2023.11.003.
Dong X., Zhang Ch., Liu Sh., Wang D. A simple and reliable semi-supervised anomaly detection network for detecting crack in stamped parts. Computers in Industry. 2025. Vol. 169. Article 104301. ISSN 0166-3615. DOI: https://doi.org/10.1016/j.compind.2025.104301
Fukatsu Y., Chen T.-T., Ogawa T., Sun F., Watanabe I., Ojima M., Ishikawa S., Adachi Y. Maximization of strength–ductility balance of dual-phase steels using generative adversarial networks and Bayesian optimization. Materials Today Communications. 2024. Vol. 41. Article № 110360. ISSN 2352-4928. DOI: https://doi.org/10.1016/j.mtcomm.2024.110360.
Wettewa S., Hou L., Zhang G. Graph Neural Networks for building and civil infrastructure operation and maintenance enhancement. Advanced Engineering Informatics. 2024. Vol. 62, Part D. Article 102868. ISSN 1474-0346. DOI: https://doi.org/10.1016/j.aei.2024.102868.
Nursyiva Irsalinda, Bakar M. A., Harun F. N., Surono S., Pratama D. A. A new hybrid approach for solving partial differential equations: Combining Physics-Informed Neural Networks with Cat-and-Mouse based Optimization. Results in Applied Mathematics. 2025. Vol. 25. Article 100539. ISSN 2590-0374. DOI: https://doi.org/10.1016/j.rinam.2025.100539.
Cao Y., Tian Z., Liu Z., Jia N., Liu X. Reducing overestimation with attentional multi-agent twin delayed deep deterministic policy gradient. Engineering Applications of Artificial Intelligence. 2025. Vol. 146. Article ID: 110352. ISSN 0952-1976. DOI: https://doi.org/10.1016/j.engappai.2025.110352.
Anbazhagan S., Mugelan R. K. Next-gen resource optimization in NB-IoT networks: Harnessing soft actor–critic reinforcement learning. Computer Networks. 2024. Vol. 252. Article № 110670. ISSN 1389-1286. DOI: https://doi.org/10.1016/j.comnet.2024.110670.
Kim S. H., Park J. S., Lee H. S., Yoo S. H., Oh K. J. Combining CNN and Grad-CAM for profitability and explainability of investment strategy: Application to the KOSPI 200 futures. Expert Systems with Applications. 2023. Vol. 225. Article ID: 120086. ISSN 0957-4174. DOI: https://doi.org/10.1016/j.eswa.2023.120086.
Hinton G., Vinyals O., Dean J. Distilling the knowledge in a neural network [Електронний ресурс]. Матеріали воркшопу “Deep Learning and Representation Learning” в рамках конференції Neural Information Processing Systems (NIPS). Монреаль, Канада, 2015. Режим доступу: https://arxiv.org/abs/1503.02531.
