文献类型：期刊文献

中文题名：基于多保真流变信息元模型的LAOS过程本构方程参数恢复

英文题名：Strategy for Parameter Recovery of Constitutive Equations in LAOS Processes Based on Multi-Fidelity Rheology-Informed Meta-Model

作者：刘虎[1];陈欣[1];张浩天[1];王雨欣[1];庞志威[1];沙金[1]

机构：[1]华东理工大学机械与动力工程学院,上海200237

年份：2025

卷号：51

期号：4

起止页码：572

中文期刊名：华东理工大学学报(自然科学版)

外文期刊名：Journal of East China University of Science and Technology

收录：;北大核心:【北大核心2023】；

语种：中文

中文关键词：大振幅振荡剪切;触变弹黏塑性本构模型;多保真;元建模;流变信息神经网络

外文关键词：large amplitude oscillatory shear;TEVP constitutive model;multi-fidelity;meta-modeling;rheology-informed neural network

摘要：为应对复杂流体在大振幅振荡剪切(LAOS)过程中触变弹黏塑性(TEVP)本构模型参数恢复的挑战,尤其在流变样本数据稀缺的情况下,提出了一种多保真流变信息元建模方法。该方法先构建了结合数据、初值和本构方程损失的流变信息神经网络作为基础子网,利用低保真合成数据进行预训练;随后引入由1×20线性层与2×20非线性层组成的多保真子网,通过加权损失函数融合低保真响应与高保真实验数据,实现元模型的增量微调。基于胶体玻璃、聚合物胶束、软凝胶和海相黏土的实验结果表明,模型预测曲线与真实数据曲线高度吻合,证明模型的有效性和鲁棒性。该方法为数据稀缺场景下的流变本构建模提供了新思路,同时展示了多保真数据融合在复杂流体参数恢复中的潜力。
Recovering parameters of constitutive equations for complex fluids under Large Amplitude Oscillatory Shear(LAOS)poses significant challenges due to the scarcity of high-quality experimental data.This study proposes a multi-fidelity rheology-informed meta-modeling framework to address this issue,focusing on the Thixotropic Elasto-Visco-Plastic(TEVP)constitutive model.The framework integrates high-fidelity experimental data(DHF)and low-fidelity synthetic data(DLF)through a Multi-Fidelity Rheology-Informed Neural Network(MF-RhiNN).The MF-RhiNN architecture consists of a base subnet with a 4×20-layer fully connected neural network,it is trained using a composite loss function that combines data loss,initial condition loss,and residuals of equations.A multi-fidelity subnet,comprising both linear and nonlinear layers,is introduced to optimize parameter recovery.Low-fidelity data,generated via numerical interpolation of initial TEVP parameters,are used for pre-training.Meanwhile,high-fidelity LAOS datasets from colloidal glass,polymer micelles,soft gels,and marine clay are employed for incremental fine-tuning.Results demonstrate that the model successfully predicts nonlinear rheological hysteresis,as evidenced by close alignment between predicted and experimental Lissajous curves.Moreover,it reveals that the nonlinear weighting factor(w)significantly impacts prediction accuracy:w=0.2 yields optimal performance,while higher weights introduce deviations,particularly under extreme strain conditions.Residual analysis confirms systematic errors in high-strain regions,highlighting areas for future improvement.Validation across diverse materials shows robust performance for colloidal glass and polymer micelles,with minor discrepancies observed in marine clay at high strains.The proposed framework effectively addresses data scarcity in rheological modeling,leveraging multi-fidelity data fusion to enhance computational efficiency and prediction reliability.This approach provides a practical solution for parameterizing complex constitutive models under LAOS.