Enhancing interactivity in structural optimisation through reinforcement learning: an application on shell structures

Abstract

This paper describes a novel approach to structural optimisation based on learning design strategies rather than searching for optimal solutions. In the proposed approach, an AI agent is trained through Reinforcement Learning (RL) to explore a 3D modelling environment and iteratively morph a flat NURBS surface into a doubly-curved shell structure. At each iteration of the 3D modelling process, the agent computes the maximum structural displacement through FEM analysis: it learns to select modelling actions through this feedback and progressively improves the performance of the input surface. Unlike current applications of RL in structural design, where the AI agent generates design options by recombining a predefined set of design variables, our approach aims to create structural forms through the interaction of a designer and an AI agent within a 3D modelling environment. An application illustrates that our agent can interpret a preliminary structural form defined by a designer and iteratively transform such a form to improve its structural performance. The trained agent can hence transform the geometry and improve the structural performance of any open surface that features a square footprint and is defined through a sequence of modelling commands. Preliminary results suggest that this AI agent can be used for the development of more interactive tools for structural design and optimisation.