Scholarlink

<p>Predicting fatigue crack growth under sparse temporal monitoring presents a significant challenge for structural prognosis. While deep learning models excel at learning spatiotemporal patterns, their purely data-driven nature can yield inaccurate predictions in critical regions especially under sparse sampling conditions. This work introduces a physics-guided SimVP (PG-SimVP) framework that integrates fundamental mechanical knowledge directly into the learning architecture. The model incorporates a differentiable physics module that constructs an analytical near-tip displacement prior based on linear elastic fracture mechanics. This prior serves as a localized guide within the training objective, steering the model toward physically realistic solutions in the crack-tip region without compromising its full-field predictive capability. Our results demonstrate that this approach effectively enhances prediction fidelity under sparse sampling conditions and improves the model’s ability to generalize to unseen loading scenarios. The framework advances the paradigm of physics-informed learning by embedding domain knowledge as an architectural component, establishing a blueprint for developing robust, generalizable physics-based machine learning models.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Physics-Guided Spatiotemporal Prediction Network for Fatigue Crack Growth

  • 期刊论文

  • 发表日期：

    2026年7月16日

查看全文
  • Licheng Jing,
  • Tienchong Chang

摘要

Predicting fatigue crack growth under sparse temporal monitoring presents a significant challenge for structural prognosis. While deep learning models excel at learning spatiotemporal patterns, their purely data-driven nature can yield inaccurate predictions in critical regions especially under sparse sampling conditions. This work introduces a physics-guided SimVP (PG-SimVP) framework that integrates fundamental mechanical knowledge directly into the learning architecture. The model incorporates a differentiable physics module that constructs an analytical near-tip displacement prior based on linear elastic fracture mechanics. This prior serves as a localized guide within the training objective, steering the model toward physically realistic solutions in the crack-tip region without compromising its full-field predictive capability. Our results demonstrate that this approach effectively enhances prediction fidelity under sparse sampling conditions and improves the model’s ability to generalize to unseen loading scenarios. The framework advances the paradigm of physics-informed learning by embedding domain knowledge as an architectural component, establishing a blueprint for developing robust, generalizable physics-based machine learning models.