预测二维泡沫通道绕障碍物流动中的塑性行为
Predicting Plasticity in Two-Dimensional Foam Channel Flow Around an Obstacle
浏览论文内容
中文总结 AI 辅助
本研究针对二维软粒子绕障碍物的受限通道流动，通过监督学习框架预测塑性活动，引入障碍物与壁面的结构描述符提升性能，指出需更优结构表示以捕捉塑性活动的详细异质性。
中文摘要 AI 辅助
我们研究二维无定形软粒子在受限通道中绕圆形障碍物流动时的塑性活动预测问题，采用基于粒子的气泡模型生成的数据集，在两种监督学习框架下构建预测任务：非仿射位移的回归任务与邻居变化事件的二分类任务。一个关键技术挑战是障碍物和受限壁面明确破坏了平移与旋转对称性，我们通过引入额外的结构描述符来编码粒子相对于这些边界的位置，以解决该问题。从简单的线性模型出发，我们逐步增加学习框架的复杂度，具体包括：对目标变量进行对数变换、纳入粒子尺寸信息、添加破坏对称性的障碍物与壁面描述符、对局部结构描述符进行粗粒化，以及使用非线性神经网络模型。我们发现，障碍物与壁面描述符对预测性能的提升最为显著。不过，此处考虑的模型主要捕捉了塑性活动在障碍物附近的整体局域化特征，未能完全复现其在单个构型中的详细异质模式。扰动分析表明，这种异质性被稳健编码在初始结构中，这意味着进一步的研究进展需要更具表达力的结构表示和机器学习架构。
英文摘要
We study the prediction of plastic activity in the confined channel flow of two-dimensional amor- phous soft particles around a circular obstacle. Using datasets generated with a particle-based bubble model, we formulate the prediction problem within two supervised-learning frameworks: re- gression of the non-affine displacement and binary classification of neighbor change events. A key technical challenge is that the obstacle and the confining walls explicitly break translational and rotational symmetries. We address this issue by introducing additional structural descriptors that encode the positions of particles relative to these boundaries. Starting from simple linear models, we systematically increase the complexity of the learning framework by considering a logarithmic transformation of the target variable, the incorporation of particle-size information, the addition of symmetry-breaking obstacle and wall descriptors, the coarse-graining of local structural descrip- tors, and nonlinear neural-network models. We find that the obstacle and wall descriptors provide the largest improvement in predictive performance. Nevertheless, the models considered here cap- ture mainly the overall localization of plastic activity near the obstacle and do not fully reproduce its detailed heterogeneous pattern in individual configurations. A perturbation analysis indicates that this heterogeneity is robustly encoded in the initial structure, suggesting that further progress requires more expressive structural representations and machine-learning architectures.
