[论文解读] Physics-based Machine Learning for Computational Fracture Mechanics
该论文提出一个基于物理的 ML 框架(phi ML),将控制方程和热力学嵌入神经网络,以建模脆性与塑性断裂,基于相场断裂数据进行训练,并在数据有限的情况下相对于简单的数据驱动模型显示出更好的泛化能力。
This study introduces a physics-based machine learning framework for modeling both brittle and ductile fractures. Unlike physics-informed neural networks, which solve partial differential equations by embedding physical laws as soft constraints in loss functions and enforcing boundary conditions via collocation points, our framework integrates physical principles, such as the governing equations and constraints, directly into the neural network architecture. This approach eliminates the dependency on problem-specific retraining for new boundary value problems, ensuring adaptability and consistency. By embedding constitutive behavior into the network's foundational design, our method represents a significant step toward unifying material modeling with machine learning for computational fracture mechanics. Specifically, a feedforward neural network is designed to embed physical laws within its architecture, ensuring thermodynamic consistency. Building on this foundation, synthetic datasets generated from finite element-based phase-field simulations are employed to train the proposed framework, focusing on capturing the homogeneous responses of brittle and ductile fractures. Detailed analyses are performed on the stored elastic energy and the dissipated work due to plasticity and fracture, demonstrating the capability of the framework to predict essential fracture features. The proposed physics-based machine learning framework overcomes the shortcomings of classical machine learning models, which rely heavily on large datasets and lack guarantees of physical principles. By leveraging its physics-integrated design, the physics-based machine learning framework demonstrates exceptional performance in predicting key properties of brittle and ductile fractures with limited training data.
研究动机与目标
- 在有限数据下,说明在非弹性条件下进行可靠断裂建模的必要性。
- 通过架构设计开发一个 guarantees 热力学一致性的 phi ML 框架。
- 在合成相场断裂数据上训练并验证 phi ML 模型,以捕捉脆性和塑性行为。
- 将 phi ML 与纯数据驱动的 FFNN 基线在准确性和外推能力方面进行比较。
- 展示这种物理整合方法的鲁棒性和数据效率。
提出的方法
- 将控制方程和热力学一致性直接嵌入到前馈神经网络架构中。
- 使用两个相关子网络:FFNN^{ε^p,d} 用于预测塑性应变和相场,FFNN^{ψ} 用于预测自由能。
- 联合训练,损失函数为 L_total = L_σ + L_ψ + L_{ε^p} + L_d + L_D,通過 D = ReLU(D_d) + ReLU(D_p) 强制耗散一致性。
- 使用修改后的 ReLU 函数对相场界限进行约束,使 d 维持在 [0,1]。
- 从有限元基础的相场仿真生成脆性和塑性断裂在均匀响应下的合成训练数据。
实验结果
研究问题
- RQ1一个基于物理的 ML 模型是否能够在有限数据下预测脆性和塑性断裂变量(应力、能量、相场、耗散)?
- RQ2物理信息感知的架构是否比纯数据驱动模型在断裂力学中具有更好的泛化性和热力学一致性?
- RQ3在数据匮乏和外推(下界/上界)条件下,phi ML 相对于纯数据驱动 FFNN 的表现如何?
- RQ4将本构行为和历史依赖变量嵌入对断裂现象的预测精度有何影响?
主要发现
- naive FFNN 在泛化方面困难,尤其在外推和数据减少时,显示出负的或接近零的 R^2 及较高的 MAPE。
- phi ML 在插值和外推下均接近完美的一致性(R^2 > 0.99),即使数据减少,对脆性断裂输出(能量、应力、相场、耗散)也表现良好。
- 对于脆性断裂,完整数据集下 phi ML 的 R^2 为 0.996–0.999,phi ML 的 MAPE ~3.18%,而 naive FFNN 为 0.894–0.998 与 14.93–58.53% 的 MAPE。
- 对于塑性断裂,表格和图表显示 phi ML 在数据有限的情况下仍保持高精度和鲁棒性,优于 naive 模型。
- 在所有脆性案例中,phi ML 框架通过将物理嵌入学习过程,展示了鲁棒性、可靠性和数据效率。
更好的研究,从现在开始
从阅读论文到最终审阅,大幅缩短您的研究时间。
无需绑定信用卡
本解读由 AI 生成,并经人工编辑审核。