Discrete Structural Optimization through a Sequential Decision Process

The aim of this award is to derive a rigorous and highly effective discrete structural optimization framework by converging structural optimization principles and sequential decision-making algorithms. Structural optimization is a design technique that is used to identify material-efficient design solutions and is widely used in many disciplines, including civil, aerospace, and mechanical engineering. Hence, new design frameworks that are capable of identifying novel and efficient solutions can improve design outcomes and be broadly beneficial by, for example, reducing the consumption of natural resources, reducing embodied carbon, improving safety and serviceability, and enhancing aesthetics. Civil structures, for example those made from steel or timber, are often constructed with standardized elements. Optimizing such structures using conventional approaches can be computationally inefficient, limited in application, or introduce approximation into the solution. By framing discrete structural optimization as a sequential decision process, that can be adeptly solved with contemporary artificial intelligence techniques, the derived framework will be particularly well suited for optimizing engineered systems constructed from standardized elements, thus leading directly to highly efficient discrete solutions and improving design outcomes. The research will be complemented by the development of an educational software application, intended for K-12 and undergraduate level students in STEM fields, that will be made publicly available to promote broad adoption and an inclusive learning opportunity about structural behavior, design, and optimization when presented as a game. The research will also provide opportunities to teach, train, and mentor students from underrepresented groups in an emerging area through outreach to various diversity programs and student organizations. The specific goal of this research is to discover the knowledge necessary to frame discrete structural optimization as a Markov Decision Process that can be adaptly solved with deep reinforcement learning techniques so as to derive a rigorous and highly effective discrete structural optimization framework. Thus, the specific research objectives of this project are to: (i) investigate how best to define the actions of the Markov Decision Process to include both topological and parametric components so as to accommodate the discrete design variables representing standardized element cross-sectional geometries; (ii) investigate and derive deep reinforcement learning solution architectures tailored for the discrete structural optimization problem; (iii) extend the framework’s applicability to the prominent volume minimization optimization problem; (iv) apply the framework to various design examples to validate and benchmark the learned policies and synthesized solutions; and (v) integrate selected design examples from the preceding objective into the development of the educational application/software where the design of truss and frame structures is presented as a game based upon sequential decision making. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.