工程结构超多维变量可微分智能优化方法研究
Research on intelligent differentiable optimization method for engineering structures with multivariables
樊健生 1杨晨 1张翀 1王琛1
作者信息
- 1. 清华大学土木工程系,北京 100084
- 折叠
摘要
参数优化问题在工程结构设计、建造、运维等阶段都极为常见.受限于有限元方法的单向分析信息流,传统优化方法主要采用启发式算法,优化效率与可优化变量数目较低,无法满足日趋复杂的工程应用需求.提出了一种基于深度学习的高效可微分结构优化方法,引入了结构体系的高保真图数据表征方法,利用智能代理模型的可微分特性,实现超多维变量的结构高效优化.不规则多层混凝土结构设计参数优化案例表明,可微分结构优化方法可在分钟级别优化数千个变量并达到令人满意的指标,材料用量比人工设计方案降低了 13％;相较于优化时间长达数星期的传统优化方法,可微分结构优化方法效率可提高10 000倍以上.可微分结构优化方法可内嵌任意工程目标函数与基于深度学习的结构智能计算模型,能够拓展应用于参数反演等多种任务,具有广泛的工程场景适应性.
Abstract
Parameter optimization problems are extremely common in all stages of engineering structure design,construction and maintenance.Constrained by the one-way analysis information flow of finite element method,the traditional optimization methods mainly use heuristic algorithms with low optimization efficiency and number of optimizable variables,which are unable to meet the demands of increasingly complex engineering applications.An efficient,differentiable structural optimization method based on deep learning was proposed,which incorporates high-fidelity graphical representations of structures,and utilizes the differentiable characteristics of intelligent agent models to achieve high efficiency of structures with multi-objective optimization of multidimensional variables.A case for the parameter optimization of irregular multi-layer concrete frame structure demonstrates that the method can optimize thousands of variables within minutes and achieve satisfactory outcomes,reducing material utilization amount by 13％compared to manual designs;compared to the weeks of optimization time required by traditional optimization method,the efficiency of the is improved by a factor of 10 000.The differentiable structural optimization method can be embedded with arbitrary engineering objective functions and deep learning-based structural intelligent computational models,and can be extended to be applied to a variety of tasks such as parameter inversion,with a wide range of adaptability to engineering scenarios.
关键词
结构优化/可微分深度学习/智能设计/工程结构/反演问题Key words
structural optimization/differentiable deep learning/intelligent design/engineering structure/inverse problem引用本文复制引用
出版年
2024
NETL
NSTL
万方数据