More Like this

1. Comparing optimization approaches in the direct displacement-based design of tall mass timber lateral systems

   Zargar, Seyed Hossein; Uarac, Patricio; Barbosa, Andre R.; Sinha, Arijit; Simpson, Barbara; van de Lindt, John W.; Brown, Nathan C. (June 2023, ASCE International Conference on Computing in Civil Engineering)

   Numerical analyses can aid design exploration, but there are several computational approaches available to consider design options. These range from “brute-force” search to optimization. However, the implementation of optimization can be challenging for the complex, time-intensive analyses required to assess seismic performance. In response to this challenge, this study tests several optimization strategies for the direct displacement-based design of a lateral force-resisting system (LFRS) using mass timber panels with U-shaped flexural plates (UFPs) and post-tensioning high-strength steel rods. The study compares two approaches: (1) a brute-force sampling of designs and data filtering to determine acceptable solutions, and (2) various automated optimization algorithms. The differential evolution algorithm was found to be the most efficient and robust approach, saving 90% of computational cost compared to brute-force sampling while producing comparable solutions. However, every optimization formulation did not return best range of design options, often requiring reformulation or hyperparameter tuning to ensure effectiveness. 

   more » « less
2. Multi-objective Optimization of Cross-Laminated Timber Shear Walls

   Shargh, A; Barbosa, Andre; Simpson, Barbara; Sinha, Arijit; van_de_Lindt, John; Brown, Nathan (January 2024, 18th World Earthquake Engineering Conference (18WCEE))

   According to a new design paradigm called Converging Design, high-level optimization objectives such as resilience and sustainability can be pursued through iterative simulation and feedback. Unlike traditional design processes that prioritize desirable seismic performance at various seismic hazard levels, the Converging Design methodology also considers the long-term ecological impact of construction and functional recovery. This methodology requires navigating competing priorities, which can be pursued through multiobjective optimization (MOO). However, computational costs and incorporating uncertainty in seismic analysis also demand that optimization frameworks use algorithms and analysis resolutions that are appropriate to the decisions being made as the design is refined. While such a framework could be applied to any material, mass timber systems are increasingly attractive as a potential sustainable solution for buildings. In this study, using a Python-based object-oriented program, an automated structural design procedure is developed to evaluate the seismic and sustainability performance of parametrically definable mass timber building configurations. Different geometric classes with Cross-Laminated Timber Rocking Walls are modeled using OpenSees and are automatically designed. Their behavior is then studied to provide insights into the relationship between structural variables and the optimization objectives. The results show a clear trade-off between Seismic Safety (the inverse of risk) and Global Warming Potential due to the construction of different design options, although the nature of this trade-off depends on the desired seismic behavior limit states. The developed software thus enables designers to efficiently explore a range of early design options for mass timber lateral systems and to achieve optimal solutions that balance seismic and sustainability performance. 

   more » « less
3. Multi-objective optimization of cross-laminated timber shear walls

   Shargh, GB; Barbosa, A; Simpson, B; Sinha, A; van_de_Lindt, JW; Brown, NC (July 2024, Proceedings of the 18th World Conference on Earthquake Engineering)

   According to a new design paradigm called Converging Design, high-level optimization objectives such as resilience and sustainability can be pursued through iterative simulation and feedback. Unlike traditional design processes that prioritize desirable seismic performance at various seismic hazard levels, the Converging Design methodology also considers the long-term ecological impact of construction and functional recovery. This methodology requires navigating competing priorities, which can be pursued through multi-objective optimization (MOO). However, computational costs and incorporating uncertainty in seismic analysis also demand that optimization frameworks use algorithms and analysis resolutions that are appropriate to the decisions being made as the design is refined. While such a framework could be applied to any material, mass timber systems are increasingly attractive as a potential sustainable solution for buildings. In this study, using a Python-based object-oriented program, an automated structural design procedure is developed to evaluate the seismic and sustainability performance of parametrically definable mass timber building configurations. Different geometric classes with Cross-Laminated Timber Rocking Walls are modeled using OpenSees and are automatically designed. Their behavior is then studied to provide insights into the relationship between structural variables and the optimization objectives. The results show a clear trade-off between Seismic Safety (the inverse of risk) and Global Warming Potential due to the construction of different design options, although the nature of this trade-off depends on the desired seismic behavior limit states. The developed software thus enables designers to efficiently explore a range of early design options for mass timber lateral systems and to achieve optimal solutions that balance seismic and sustainability performance. 

   more » « less
4. Finding Shortest Reconfiguration Sequences for Modular Robots (Media Exposition)

   https://doi.org/10.4230/lipics.socg.2025.85  

   A_Akitaya, Hugo; Clements, Andrew; Downey, Sam; Eisenbies, Jonathan; Samanta, Soham; Shahrouzi, Gabriel; Stock, Frederick (June 2025, Schloss Dagstuhl – Leibniz-Zentrum für Informatik)

   Aichholzer, Oswin; Wang, Haitao (Ed.)

   This paper introduces a set of tools built to help researchers design algorithms for modular robots. These tools can brute force solutions to specific reconfigurations, visualize movements of modular robots, and can be used to design specific configurations of robots. Multiple models of modular robots are supported, and can be added by users. 

   more » « less
5. Analysis of Deep Reinforcement Learning for Control Co-Design

   Gilik, Aysenur; Bayrak, Alparslan Emrah (August 2026, ASME 2026 International Design Engineering Technical Conferences & Computers and Information in Engineering Conference)

   Concurrent solution of coupled physical and control system design problems, also known as control co-design, is the key for innovation in active or smart systems. Existing control co-design approaches in the optimization literature are mostly model-based and rely on a prediction of the system response to a control action. A model-free approach such as reinforcement learning can explore physical and control system design candidates using trial-and-error and identify promising solutions for complex smart systems. Leveraging the advances in reinforcement learning, we discuss the use of Q-learning and double deep Q-networks as model-free approaches in control co-design using a bi-level formulation. We compare the two approaches in terms of performance, computational cost and scalability. We illustrate the application of the proposed approach on a modified version of an open-source benchmark problem, the MountainCar challenge by Open AI. The control system design problem involves finding a driving strategy for a car to climb to the top of a mountain and the physical system design problem aims to find the best final drive ratio. We benchmark the control co-design results against those from a brute-force solution that trains a separate control policy for all possible discrete physical system design variable values to verify the convergence to system optimality. We discuss the generalizability of the present bi-level formulation by adopting other reinforcement learning methods beyond Q-learning or double deep Q-networks. 

   more » « less