Frydrych Karol, Tomczak Maciej, Papanikolaou Stefanos
NOMATEN Centre of Excellence, National Centre for Nuclear Research, Sołtana 7, 05-400 Otwock, Poland.
Institute of Fundamental Technological Research, Polish Academy of Sciences, Pawińskiego 5b, 02-106 Warsaw, Poland.
Materials (Basel). 2024 Jul 9;17(14):3397. doi: 10.3390/ma17143397.
This paper describes an application of a machine learning approach for parameter optimization. The method is demonstrated for the elasto-viscoplastic model with both isotropic and kinematic hardening. It is shown that the proposed method based on long short-term memory networks allowed a reasonable agreement of stress-strain curves to be obtained for cyclic deformation in a low-cycle fatigue regime. The main advantage of the proposed approach over traditional optimization schemes lies in the possibility of obtaining parameters for a new material without the necessity of conducting any further optimizations. As the power and robustness of the developed method was demonstrated for very challenging problems (cyclic deformation, crystal plasticity, self-consistent model and isotropic and kinematic hardening), it is directly applicable to other experiments and models.
本文描述了一种用于参数优化的机器学习方法的应用。该方法在具有各向同性和运动硬化的弹粘塑性模型中得到了验证。结果表明，基于长短期记忆网络的所提方法能够在低周疲劳状态下的循环变形中获得应力-应变曲线的合理拟合。所提方法相对于传统优化方案的主要优势在于，无需进行任何进一步优化即可获得新材料的参数。由于所开发方法的能力和稳健性在极具挑战性的问题（循环变形、晶体塑性、自洽模型以及各向同性和运动硬化）中得到了验证，因此它可直接应用于其他实验和模型。
