Deep Reinforcement Learning for Engineering Design through Topology Optimization of Elementally Discretized Design Domains

Nathan Brown, Anthony Garland, Georges Fadel, Gang Li

Keywords: Reinforcement Learning, Machine Learning, Topology Optimization, Engineering Design

TL;DR: A deep reinforcement learning agent was successfully trained to design 2D topologies that satisfy topology optimization objectives.

Abstract: Machine learning (ML) can extract patterns in design-relevant data to detect trends or make predictions that may not be inherently visible to a human designer. However, most ML-based engineering design tools rely on supervised learning, which requires the prefabrication of design domain data that may be challenging to derive or inherently biased by a human designer. This work addresses these limitations by investigating the implementation of reinforcement learning (RL), a unique subset of ML that learns through accumulating past experiences in an interactive environment, into the engineering design problem of topology optimization. RL offers the design freedom and complexity of a human designer while maintaining the computational efficiency of more common machine learning paradigms. An RL environment was formatted to allow a deep RL agent to design 2D elementally discretized topologies based on a multi-objective reward function. After training, the agent was tested using progressive refinement on a variety of common load cases to validate the design capabilities and generalization of the agent. The results, which proved to be comparable to a traditional gradient-based topology optimization solver, show that a deep RL agent can learn generalized design strategies to satisfy multi-objective design tasks and, therefore, shows promise as a design tool for arbitrarily complex design problems across many design domains.