Machine-learning-guided control of a pulsed cold atmospheric plasma jet reveals a voltage-dependent discharge transition relevant to diabetic wound healing.

Cold atmospheric plasma (CAP) technology is promising for biomedical treatment, but its translation is limited by the difficulty of identifying operating conditions that simultaneously preserve non-thermal discharge stability, enhance reactive-species generation, and yield reproducible biological responses. Here, a pulsed dielectric-barrier-discharge CAP jet operated from 5 to 45 kV was investigated by combining electrical diagnostics, optical emission spectroscopy, and machine learning to define a treatment-relevant operating window for preclinical diabetic wound repair. Voltage and current waveforms, electron temperature, electron density, plume length, and output voltage were quantified across the operating range, while gas temperature was monitored to confirm non-thermal operation. The combined diagnostics indicated a voltage-dependent transition region around 29-31 kV, marked by nonlinear current scaling, a change in plume-length behavior, and an increase in electron density while gas temperature remained within a non-thermal range. Multilayer perceptron (MLP) and radial basis function (RBF) networks were then trained to predict plasma-state variables and wound-size trajectories under lower- and higher-voltage treatment windows. Across most tasks, MLP provided more consistent performance than RBF, particularly near the transition region, indicating stronger capacity to represent nonlinear discharge behavior and its association with downstream biological response. In a diabetic rat wound model, the higher-voltage and larger stand-off condition produced the most favorable endpoint wound-size reduction while preserving non-thermal operation. These results establish a machine-learning-assisted framework for mapping diagnostically defined CAP operating windows and support future development of closed-loop CAP systems for reproducible biomedical use.