Additive manufacturing offers exceptional flexibility in producing complex and customized structures. However, achieving high precision requires more than executing a predefined process. Most systems rely on static workflows in which a 3D model is sliced and converted into G-code that dictates the heat source path, power, and timing. These static instructions do not account for the underlying physics of the printing process, such as heat transfer, material behavior, and geometric complexity. As a result, defects like balling, warping, and delamination may occur. In processes like powder bed fusion (PBF) and vat photopolymerization (VAT), one of the key factors influencing these outcomes is the toolpath strategy. The scanning sequence, direction, and spacing have a direct impact on heat accumulation, residual stress, and thermal consistency. These thermal effects, in turn, influence mechanical integrity and surface quality [1]. Traditional fixed-path approaches often lead to localized overheating or inconsistent thermal conditions that cannot be fully corrected by process calibration alone. To address these limitations, adaptive control methods are being explored to improve path planning in real time. One promising direction is reinforcement learning (RL), a machine learning approach in which agents learn decision-making policies by interacting with an environment to maximize a reward. Prior studies by Qin et al. [2] and Mozaffar et al. [3] demonstrate the potential of RL to generate optimized toolpaths using simulated feedback, although their methods are based on simplified physical assumptions and do not incorporate physics-based modeling. In this work, we aim to close this gap by coupling RL with detailed numerical simulations. Thermal gradients and stress distributions are key to defect formation in additive manufacturing processes like PBF and VAT. To account for these effects, the learning process takes place in an environment that includes a finite element model capable of accurately simulating these phenomena. To optimize for an adaptive toolpath within this environment, we apply Deep Q-Networks (DQN) and Proximal Policy Optimization (PPO). By explicitly modeling the thermally driven behavior of these systems, we create more realistic environments for learning-based decision making. As a proof of concept, we demonstrate how the method can optimize printing speed for different geometries.