Determination of Electron Velocity Distribution Functions via Neural Network-based Smoothing

In partially magnetized E×B sources, spatial gradients in plasma parameters induce azimuthal drift motion of electrons, making anisotropic behavior in the electron velocity phase inevitable and necessitating Mach probe measurements of the electron velocity distribution function (EVDF). Measured EVDFs exhibit non‐Maxwellian, anisotropic features; accordingly, plasma parameters must be extracted by integrating the EVDF over distinct momentum intervals. However, fitting the ion current and extracting the electron current remain challenging because sheath expansion becomes highly nonlinear and unpredictable. In addition, inevitable fluctuations in probe currents in magnetized plasmas complicate measurements, and a paucity of data points near zero velocity further degrades plasma‐parameter calculations. Here, we introduce a multilayer perceptron (MLP) neural network with ensemble averaging to reconstruct the EVDF from two‐sided planar probe data in partially magnetized, low‐temperature plasmas. The network captures the overall shape of the current–voltage characteristic for robust, adaptive smoothing without manual parameter adjustment. By fitting ion current at low bias and performing lossless interpolation near the plasma potential, this approach reduces noise and fluctuations in derivative data across the entire electron energy range, improving the accuracy and reliability of EVDF extraction under conditions in which conventional smoothing techniques fail.