Kenny Angus, Ray Tapabrata, Singh Hemant
School of Engineering and Technology, University of New South Wales, Canberra, ACT, 2600, Australia.
Sci Rep. 2024 Apr 3;14(1):7858. doi: 10.1038/s41598-024-57468-2.
In engineering design, there often exist multiple conceptual solutions to a given problem. Concept design and selection is the first phase of the design process that is estimated to affect up to 70% of the life cycle cost of a product. Currently, optimization methods are rarely used in this phase, since standard optimization methods inherently assume a fixed (given) concept; and undertaking a full-fledged optimization for each possible concept is untenable. In this paper, we aim to address this gap by developing a framework that searches for optimum solutions efficiently across multiple concepts, where each concept may be defined using a different number, or type, of variables (continuous, binary, discrete, categorical etc.). The proposed approach makes progressive data-driven decisions regarding which concept(s) and corresponding solution(s) should be evaluated over the course of search, so as to minimize the computational budget spent on less promising concepts, as well as ensuring that the search does not prematurely converge to a non-optimal concept. This is achieved through the use of a tree-structured Parzen estimator (TPE) based sampler in addition to Gaussian process (GP), and random forest (RF) regressors. Aside from extending the use of GP and RF to search across multiple concepts, this study highlights the previously unexplored benefits of TPE for design optimization. The performance of the approach is demonstrated using diverse case studies, including design of a cantilever beam, coronary stents, and lattice structures using a limited computational budget. We believe this contribution fills an important gap and capitalizes on the developments in the machine learning domain to support designers involved in concept-based design.
在工程设计中，对于给定问题通常存在多种概念性解决方案。概念设计与选择是设计过程的第一阶段，据估计该阶段会对产品生命周期成本产生高达70%的影响。目前，优化方法很少用于此阶段，因为标准优化方法本质上假定概念是固定（给定）的；而对每个可能的概念进行全面优化是不可行的。在本文中，我们旨在通过开发一个框架来填补这一空白，该框架能够在多个概念中高效地搜索最优解，其中每个概念可能使用不同数量或类型的变量（连续、二元、离散、分类等）来定义。所提出的方法在搜索过程中就应该评估哪些概念和相应的解决方案做出渐进式数据驱动决策，以便将花费在前景不太乐观的概念上的计算预算降至最低，同时确保搜索不会过早收敛到非最优概念。这是通过除了使用高斯过程（GP）和随机森林（RF）回归器之外，还使用基于树结构帕曾估计器（TPE）的采样器来实现的。除了将GP和RF的应用扩展到跨多个概念的搜索之外，本研究还突出了TPE在设计优化方面此前未被探索的优势。使用多种案例研究展示了该方法的性能，包括使用有限计算预算设计悬臂梁、冠状动脉支架和晶格结构。我们相信这一贡献填补了一个重要空白，并利用机器学习领域的发展来支持参与基于概念设计的设计师。
