文献类型：期刊文献

中文题名：基于物理信息深度学习的高分子流变学研究:挑战、方法与应用

英文题名：Physics-Informed Deep Learning for Polymer Rheology:Investigating Challenges,Methodologies,and Application

作者：祁纪浩[1];陈欣[1];庞志威[1];林增[1];王超元[1];刘虎[1];沙金[1];白志山[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：56

期号：11

起止页码：1910

中文期刊名：高分子学报

外文期刊名：Acta Polymerica Sinica

收录：;北大核心:【北大核心2023】；

基金：国家杰出青年科学基金(基金号22225804);国家自然科学基金(基金号22408101);上海市自然科学基金(基金号25ZR1401085);贵州省科技支撑计划(项目号黔科合支撑[2025]一般091)资助.

语种：中文

中文关键词：物理信息深度学习;高分子流变学;建模与仿真;智能预测与优化

外文关键词：Physics-informed deep learning;Polymer rheology;Modeling and simulation;Intelligent prediction and optimization

摘要：高分子流变学旨在理解材料从微观到宏观的流动与变形特性,但传统建模方法在处理复杂非线性、多尺度模拟及数据解析方面长期面临挑战.物理信息深度学习(PIDL)通过将流变学物理定律嵌入深度神经网络,减少对大量数据的依赖,提高模型泛化能力和物理合理性,为解决高分子流变学面临的挑战提供新的途径.通过探讨PIDL的基本原理,包括其模型架构、物理信息损失函数设计(如MSE、MAE及稀疏/残差损失)和损失函数优化方法(如梯度下降、自适应权重调整),并分析PIDL在高分子流变学中的具体应用,涵盖了概率模型、生成式模型、神经算子、图神经网络和强化学习等多种模型架构,展示其在参数优化、流变特性预测、逆问题求解及数字孪生集成等方面的潜力.尽管PIDL展现出显著优势,但仍面临数据获取成本高、模型可解释性弱、多尺度求解精度不足及鲁棒性等问题.未来发展方向包括利用多保真技术、物理信息增强数据、优化算法、结合数字孪生、推断因果关系以及应用元学习和可解释人工智能(XAI)来提升模型性能和实用性.
Polymer rheology aims to understand the multiscale flow and deformation of macromolecular materials,yet conventional constitutive approaches have long struggled with pronounced nonlinearities,cross-scale coupling and sparse,noisy data inversion.Physics-informed deep learning(PIDL)embeds rheological conservation laws and constitutive relations directly into deep neural networks,substantially reducing the demand for large labelled datasets while enhancing model generalization and physical interpretability.This review systematically outlines the foundational principles of PIDL,covering network architectures,physics-informed loss functions such as MSE,MAE,sparse and residual variants,and optimization strategies including gradient descent and adaptive weighting.We further survey five PIDL paradigms—probabilistic models,generative models,neural operators,graph neural networks and reinforcement learning—and demonstrate their potential for parameter identification,rheological property prediction,inverse problem solving and digital-twin integration.Despite these advantages,PIDL still contends with high data acquisition costs,limited interpretability,insufficient accuracy across scales and uncertain robustness under extreme conditions.Future research should exploit multi-fidelity techniques,physics-augmented data generation,advanced optimization,causal inference and explainable AI to deliver trustworthy,industrially viable rheological models.