Machine-Learning Data Processing for Accurate Langmuir Probe Diagnostics

Langmuir probes have been a cornerstone of plasma diagnostics. However, the sensitivity of probe signals and their dependence on human interpretation compromise consistency, so even experts can reach different conclusions from the same I–V dataset. In particular, in collisional plasmas, accurate subtraction of the ion current with a conventional model is trivial; second-order differentiation amplifies noise from residual ion-current error, producing distortions in the extracted electron energy distribution and derived electron density (ne) and temperature (Te). Here, we introduce a strategy that combines physics-based preprocessing with a machine-learning regressor to extract plasma parameters from raw I–V data. The model is trained on synthetic probe characteristics spanning realistic collisionalities, sheath potentials, and noise levels, learning the nonlinear mapping from measured current to EEPFs. This approach was validated with RF-compensated Langmuir probe data collected in a 13.56 MHz inductively coupled argon plasma at 300 mTorr and 500 mTorr. Compared with conventional Druyvesteyn analysis, the machine-learning approach corrected ne underestimation and returned smooth, non-negative EEPFs without manual intervention. Our results suggest that AI signal processing can restore the diagnostic value of Langmuir probes in collisional plasmas, enabling real-time, operator-independent monitoring in both laboratory research and industrial plasma processing.