Structured Comparison of Multi-Layer Perceptron Design Elements for Accurate Smoothing of Electron Energy Probability Functions

Accurate interpretation of Langmuir probe current–voltage (I–V) characteristics is essential for reliable extraction of the electron energy probability function (EEPF), especially in non-equilibrium plasma environments. While neural network–based smoothing using multi-layer perceptron (MLP) has shown significant advantages over conventional methods in reconstructing smooth and physically meaningful second derivatives, the impact of internal MLP configurations has not yet been fully explored. In this study, we aim to refine EEPF accuracy further by systematically evaluating how various MLP design elements influence I–V curve analysis. The investigation is structured into six key categories: data normalization, activation functions, network architecture, regularization and optimization techniques, ensemble strategies, and input segmentation schemes. For each category, representative techniques are applied and compared using metrics such as second derivative smoothness, parameter stability, noise suppression, and overfitting resistance. By identifying optimal design strategies within the MLP framework, this work offers practical guidance for improving Langmuir probe diagnostics and contributes to more robust and interpretable extraction of EEPFs.