Abstract Predicting history-dependent materials’ responses is crucial, as path-dependent behavior appears while characterizing or geometrically designing many materials (e.g., metallic and polymeric cellular materials), and it takes place in manufacturing and processing of many materials (e.g., metal solidification). Such phenomena can be computationally intensive and challenging when numerical schemes such as the finite element method are used. Here, we have applied a variety of sequence learning models to almost instantly predict the history-dependent responses (stresses and energy) of a class of cellular materials as well as the multiphysics problem of steel solidification with multiple thermo-viscoplasticity constitutive models accounting for substantial temperature, time, and path dependencies, and phase transformation. We have shown the gated recurrent unit (GRU) as well as the temporal convolutional network (TCN), can both accurately learn and almost instantly predict these irreversible, and history- and time-dependent phenomena, while TCN is more computationally efficient during the training process. This work may open the door for the broader adoption of data-driven models in similar computationally challenging constitutive models in plasticity and inelasticity.

中文翻译：

---

塑性和热粘塑性的深度学习

摘要 预测历史相关材料的响应至关重要，因为在表征或几何设计许多材料（例如金属和聚合物蜂窝材料）时会出现路径相关行为，并且它发生在许多材料的制造和加工中（例如，金属凝固）。当使用诸如有限元方法之类的数值方案时，此类现象可能是计算密集型且具有挑战性的。在这里，我们应用了各种序列学习模型来几乎立即预测一类细胞材料的历史相关响应（应力和能量）以及钢凝固的多物理场问题，其中多个热粘塑性本构模型解释了大量温度、时间和路径相关性以及相变。我们已经展示了门控循环单元 (GRU) 以及时间卷积网络 (TCN)，可以准确地学习并几乎立即预测这些不可逆的、与历史和时间相关的现象，而 TCN 在训练期间的计算效率更高过程。这项工作可能为在类似的具有计算挑战性的可塑性和非弹性本构模型中更广泛地采用数据驱动模型打开大门。