Closed-loop reinforcement learning control of AM process parameters using Digital Twin feedback for defect mitigation and print correction

Additive Manufacturing (AM) offers significant design flexibility but suffers from defects such as porosity, lack of fusion and warping due to complex thermo-physical interactions during fabrication. Conventional controllers, including PID-based approaches, struggle with these challenges because observability between surface measurements and subsurface thermal states is inherently limited. This paper presents a Twin-in-the-Loop system combining high-fidelity Digital Twins (DTs) with Reinforcement Learning (RL) for autonomous, closed-loop AM process control enabling real-time defect mitigation. We introduce a Cyber-Physical-Digital (CPD) architecture with three layers: (i) a multi-modal sensing layer using infrared thermography, pyrometry, acoustic emission (AE) and high-speed vision; (ii) a Digital Twin layer driven by Fourier Neural Operators (FNOs) predicting 3D thermal fields with 12.4 ms inference and 3.8% MAPE; (iii) a Soft Actor-Critic (SAC) RL controller adjusting laser power, scan speed and feed rate at 50 Hz. Acoustic-emission integration enables detection of crack-initiation events invisible to thermal and optical sensors. Ti-6Al-4V Laser Powder Bed Fusion experiments validated by X-ray Computed Tomography achieved 99.79% relative density, substantially surpassing open-loop baselines. The system deploys on edge devices such as NVIDIA Jetson AGX Orin, enabling real-time, autonomous defect-free AM production without cloud infrastructure.