Reinforcement Learning-Based Path Planning for Additive Manufacturing of Homogenised Grid Structures

Abstract

This study provides new insight into the mechanisms linking shrinkage to long-term degrad This thesis presents a methodology for designing and manufacturing highly tunable Repetitive volume element (RVE) with tunable isotropy and mechanical properties through homogenisation reinforcement learning-guided toolpath planning for material extrusion additive manufacturing.
Background and Motivation FDM is used to create highly adaptable objects with geometrically complex shapes and features. High anisotropic behaviour is measured in conventional methods of FDM, and by optimising the mechanical properties and a toolpath of a given RVE, the result becomes highly variable. Scaling the material properties using a combination of homogenisation and RL-based toolpath planning helps extend the uses of AM.
Methodology The approach consists of three main stages. First, 2D voxel-based grids with controlled infill densities of 63% to 93% are generated to create permeable structures. Homogenisation techniques are applied to determine effective mechanical properties, e.g. Young’s moduli, shear moduli and Poisson’s ratios, across multiple configurations varying in voxel size, 0.1-0.4 mm, and void distribution. Second, optimal grid designs are selected using sequential optimisation with a multi-objective function balancing mechanical stiffness, isotropy, shear performance, and geometric connectivity. Third, a Double Deep Q-Network (DDQN) reinforcement learning agent is trained on 50 diverse grids to generate void-aware toolpaths for a 0.4 mm nozzle. The agent learns to maximise material coverage while avoiding collisions with voids, minimising redundant motions, and ensuring printability.
Results The trained RL agent achieved 89.4% material coverage on test grids with infill ratios of 73%, demonstrating effective void avoidance and path continuity. The optimised grid design exhibited an improvement in isotropy compared to conventional raster patterns, with 1040% reduction in mechanical anisotropy. RVEs were fabricated via MEX and mechanically tested, validating the computational predictions within close accuracy based on the isotropic behaviour, yet tensile tests show lower Young’s moduli than expected from the computational model.
Significance and Future Work This work demonstrates that reinforcement learning can effectively optimise toolpath planning for complex, defect-containing structures, enabling the fabrication of FDMproduced objects with improved mechanical isotropy. Future work should extend the methodology to 3D homogenisation to eliminate plane-stress/plane-strain approximations, incorporate multi-directional AM for true 3D toolpaths, and explore non-planar geometries that more closely replicate native tissue architecture. These advances would optimise the tunable properties already existing in AM.ation in AAS and demonstrates that combining sodium silicate activation, gypsum addition, and early fog curing offers a practical route to durable, shrinkage-resistant, and sustainable AAS concretes suitable for structural applications.