发表机构
Northwestern University（西北大学）
机构由 AI 辅助整理，请以论文原文为准。AI 中文总结
本研究在广义标准材料框架内提出统一的可学习内聚力模型，用凸与单调神经网络表示能量和损伤抗力，无需特定定律即可从数据学习多种复杂内聚行为。
AI 中文摘要
内聚力模型被广泛用于描述断裂和界面失效，然而大多数公式预先规定了针对特定问题的解析牵引-分离定律以及卸载和再加载的唯象规则，导致针对不同内聚行为需要专门的模型。本工作在广义标准材料框架内发展了一种统一的可学习内聚公式，其中响应由学习得到的本构函数生成，而底层热力学结构保持不变。表面自由能被分解为主动和接触贡献，不可逆损伤演化由学习得到的依赖于模式的损伤抗力控制。主动能量由输入凸神经网络表示，而逆损伤抗力由单调神经网络表示。凸性、单调性、归一化和损伤不可逆性被直接纳入本构表示。逆抗力的直接参数化产生了显式的损伤更新，并避免了本构评估过程中的局部非线性反演。材料点研究表明，该公式能够表示定性不同的内聚响应，包括平台、延长的软化尾、不规则软化、非线性卸载、不同的I型和II型行为，以及若干经典混合模式内聚定律。因此，该框架用单一的热力学结构化表示取代了特定定律的模型构建，该表示能够从数据中学习广泛功能复杂性的内聚响应。
英文摘要
Cohesive zone models are widely used to describe fracture and interfacial failure, yet most formulations prescribe problem-specific analytical traction-separation laws together with phenomenological rules for unloading and reloading, leading to specialized models for different cohesive behaviors. This work develops a unified learnable cohesive formulation within the generalized standard materials framework, in which the response is generated from learned constitutive functions while the underlying thermodynamic structure remains fixed. The surface free energy is decomposed into active and contact contributions, and irreversible damage evolution is governed by a learned mode-dependent damage resistance. The active energy is represented by an input-convex neural network, while the inverse damage resistance is represented by a monotone neural network. Convexity, monotonicity, normalization, and damage irreversibility are incorporated directly into the constitutive representation. Direct parameterization of the inverse resistance yields an explicit damage update and avoids local nonlinear inversion during constitutive evaluation. Material-point studies show that the formulation can represent qualitatively distinct cohesive responses, including plateaus, extended softening tails, irregular softening, nonlinear unloading, distinct Mode I and Mode II behaviors, and several classical mixed-mode cohesive laws. The framework therefore replaces law-specific model construction with a single thermodynamically structured representation capable of learning cohesive responses of broad functional complexity from data.
