王琛, 何有泉, 宋凌寒, 樊健生. 理论辅助的弹塑性本构关系小样本深度学习[J]. 工程力学, 2023, 40(9): 29-36. DOI: 10.6052/j.issn.1000-4750.2021.12.1012
引用本文: 王琛, 何有泉, 宋凌寒, 樊健生. 理论辅助的弹塑性本构关系小样本深度学习[J]. 工程力学, 2023, 40(9): 29-36. DOI: 10.6052/j.issn.1000-4750.2021.12.1012
WANG Chen, HE You-quan, SONG Ling-han, FAN Jian-sheng. A THEORY-AIDED FEW-SHOT DEEP LEARNING ALGORITHM FOR ELASTOPLASTIC CONSTITUTIVE RELATIONSHIPS[J]. Engineering Mechanics, 2023, 40(9): 29-36. DOI: 10.6052/j.issn.1000-4750.2021.12.1012
Citation: WANG Chen, HE You-quan, SONG Ling-han, FAN Jian-sheng. A THEORY-AIDED FEW-SHOT DEEP LEARNING ALGORITHM FOR ELASTOPLASTIC CONSTITUTIVE RELATIONSHIPS[J]. Engineering Mechanics, 2023, 40(9): 29-36. DOI: 10.6052/j.issn.1000-4750.2021.12.1012

理论辅助的弹塑性本构关系小样本深度学习

A THEORY-AIDED FEW-SHOT DEEP LEARNING ALGORITHM FOR ELASTOPLASTIC CONSTITUTIVE RELATIONSHIPS

  • 摘要: 该文提出了一种引入经典弹塑性力学理论知识作为辅助驱动的小样本深度学习算法,适用于土木工程任意材料弹塑性本构关系,能够有效缓解大规模深度学习模型实际应用时常见的数据量匮乏瓶颈。该文简要概述了通用的经典弹塑性力学框架;在此基础上详细阐释了将弹塑性力学方程引入到常规深度学习模型中的方法与流程,该过程无需关心底层理论本构模型的具体形式与传统的复杂数值实现,保留了数据驱动技术简单、直接、高效的优点;为缓解优化目标复杂化所导致的训练不收敛问题,提出了一种与理论辅助驱动相适应的训练策略“过拟合-修正法”,能够稳定并加速收敛过程;基于结构钢材精细弹塑性本构模型开展了数值试验,验证了理论辅助的小样本学习算法的有效性,能够实现大规模深度学习模型在少量训练样本情形下获得优异的泛化性,相较纯粹数据驱动模型准确性提升38.9%。该文采用的理论辅助思想具有可借鉴性,后续可应用于结构层次的深度学习代理模型研究,促进未来更为先进、大型的智能算法落地土木工程计算领域。

     

    Abstract: Proposed is a theory-aided few-shot learning algorithm applicable to any material elastoplastic relationships utilized in civil engineering. The new algorithm can effectively alleviate the paucity of data in real applications when applying large-scale deep learning models. The framework of the classical elastoplasticity theory is briefly recapped. On this basis, the paper elaborates on how to incorporate the elastoplastic equations into the ordinary deep learning models, wherein the entire processes require neither the concrete form of the underlying models nor its complicated numerical implementation, thus maintaining the simple, direct, and efficient advantages of data-driven techniques. To address the divergence issue caused by the complicated optimization target, a novel training strategy named overfitting-correction method is invented, which is able to stabilize the training and accelerate the convergence. A numerical experiment is performed upon a sophisticated elastoplastic constitutive model of structural steel. The results demonstrate that the theory-aided few-shot learning algorithm succeeds in realizing excellent generalization performance of a large-scale deep learning model on small training datasets. Quantitatively, the new algorithm overwhelms the pure data-driven model by 38.9%. The philosophy of theory-aided driven mode proposed can extend to the structural level in the future, facilitating the introduction of more advanced intelligent technologies into civil engineering.

     

/

返回文章
返回