ParamNet: A Physics‐Guided Deep Learning Framework for Intelligent Self‐Inversion of Vacuum Optical Levitation Systems

Optical tweezers have become a pivotal tool in fundamental physics, precision metrology, and macroscopic quantum phenomena, enabling precise manipulation of microparticles. Accurate calibration of optical traps, specifically the trap stiffness and viscous damping coefficient , is essential for precise force measurements and particle manipulation in quantitative experiments. In this article, we introduce ParamNet, a physics‐guided deep learning framework for parameter inversion in vacuum optical tweezers, combining data‐driven feature learning with model‐based constraints for fast, accurate calibration. ParamNet employs a dual‐branch time–frequency architecture that extracts complementary features from time‐domain trajectories and frequency‐domain representations, facilitating efficient information fusion and improved calibration accuracy. The network achieves a mean absolute percentage error of approximately 3% for and 7% for under simulated scenarios, using short windows of 100 samples (0.01 s), outperforming traditional methods in simulation and experiments. Validation results show that ParamNet provides a compact and transferable framework for intelligent self‐inversion of vacuum optical tweezers. It reduces the reliance on long‐duration or high‐rate calibration data and shows potential for digital‐twin‐style monitoring and feedback‐oriented adaptive control of levitated systems.