Computer Methods in Materials Science
Loading...
ISSN 2720-4081
e-ISSN: 2720-3948
Issue Date
2026
Volume
Vol. 26
Number
No. 1
Description
Journal Volume
Computer Methods in Materials Science
Vol. 26 (2026)
Projects
Pages
Articles
A Digital Twin for temperature prediction in the laser hardening process of NC10 steel
(Wydawnictwa AGH, 2026) Lacki, Piotr; Derlatka, Anna; Lacki, Michał; Lachs, Kuba
In this study, Artificial Neural Networks (ANN) were created to develop a Digital Twin (DT) for temperature prediction in the laser hardening process of NC10 steel. The ANN were trained to predict temperature on the top layer during the laser hardening process of NC10 steel samples with different thicknesses and with various laser power and laser scanning speeds. The prediction developed during the project work was based on a parametric numerical model of the laser hardening process for a sample of NC10 steel, using the Finite Element Method (FEM) within the ADINA software. Numerical simulations enabled a detailed analysis of the temperature produced on the surface of each sample, as well as a visualization of the structural changes made to the sample according to the laser hardening process. It is crucial to create data that reflects reality as closely as possible to assess the best setting for each process. A well created DT allows to make automatically important changes along laser hardening process. To obtain a set of the most efficient parameters for the desired result, Genetic Algorithms (GA) were integrated with the developed ANN. As a result, the authors developed an effective and efficient tool to predict the temperature produced along the laser hardening process.
Geometric optimization of two-stage stamping dies for ultra-thin titanium bipolar plates using Sequential Physics-Informed Neural Networks
(Wydawnictwa AGH, 2026) Ke, Zijie; Huang, Yiwen; Guo, Ziqiang; Xiao, Yao; Hou, Zeran; Min, Junying
Bipolar plates are critical core components in proton exchange membrane fuel cells (PEMFCs). Titanium-based materials are highly favored due to their excellent corrosion resistance and high specific strength. However, the plates often experience severe local thinning and poor consistency in forming dimensions during the two-stage stamping process. Although traditional finite element method (FEM) optimization can mitigate these defects, it comes with high computational costs and time consumption. This study proposes a die design optimization framework based on the Sequential Physics-Informed Neural Network (S-PINN). Unlike traditional single-layer neural network models, S-PINN adopts a sequential architecture that effectively maps the two-stage forming process of the plates. This architecture can explicitly predict the evolution of forming quality from the pre-forming stage to the final stage. By embedding the core physical laws of plastic deformation into the network loss function, the S-PINN model effectively predicts the complex nonlinear relationship between mold geometry and forming quality, while ensuring physical consistency. Experimental and simulation results show that the S-PINN model’s prediction accuracy for dimensional consistency (DC) is 73.8% higher than that of the PINN model and 33.9% higher than that of the S-ANN model. Compared with traditional modeling methods, the S-PINN-optimized die design can reduce the thinning rate and improve channel dimensional consistency.
Artificial intelligence-enhanced algebraic multigrid for 3D finite element simulations
(Wydawnictwa AGH, 2026) Goik, Damian; Banaś, Krzysztof
The paper presents preliminary investigations into a strategy for solving linear systems resulting from 3D finite element simulations, based on the algebraic multigrid (AMG) method, enhanced using artificial intelligence techniques. In particular, we adapt to 3D problems the algorithm presented in Luz et al. (2020) for using a graph neural network to create the prolongation and restriction operators in a way that will improve convergence. The process of training the network proceeds on the basis of a set of system matrices obtainedfor tasks much smaller in scale than the target problem to be solved. Learning is aimed at decreasing the spectral radius of the error propagation matrix after applying modified prolongation and restriction. We describe some implementation details of the solver developed based on the presented strategy and show several numerical examples of its application for medium-sized problems.