A Deep Reinforcement Learning Approach for Real-World 3D Facility Layout Problems

The automatic development and optimization of facility layouts is a costly and time-consuming stage in the design phase of a manufacturing or process plant. Both the planning efforts and the quality of the final layout significantly impact the overall project costs. Despite advances in automation and artificial intelligence, the planning process still requires a high degree of manual effort, especially when dealing with real-world environments. Recent deep reinforcement learning (DRL) approaches show promising results in automatically generating optimized layouts in discrete 2D environments. However, most real-world use cases require placing plant components in continuous 3D brownfield environments, which demand a high degree of positional detail unattainable with 2D approximations. To address this research gap, a DRL method is proposed to optimize facility layout problems (FLPs) by using real-world collision shapes of brownfield and facility components in a physics-based continuous 3D environment. The method is evaluated using a 3D model of an electrolyzer plant, where its main components are placed in a brownfield while minimizing material handling costs (MHC). The results indicate that the DRL method can optimize MHC by placing components within the brownfield with high accuracy, overcoming the shortcomings of simplified discrete approaches.