HyPhysML: analysis code for physics-informed flashover voltage prediction
Year: 2026
Type: software
Abstract: Analysis code accompanying the manuscript "Physics-informed hybrid machinelearning predicts contamination-induced flashover voltage of outdoorinsulators" (M. Burukanli), submitted to Scientific Reports. The repository contains the complete pipeline in a single executable script:feature engineering over a 25-dimensional physics-informed feature space;Optuna-TPE hyperparameter optimisation on an independent holdout; anout-of-fold stacking ensemble of seven base learners with a ridgemeta-learner; a validation layer that checks the fitted Obenaus log-linearcoefficients against their expected signs; Friedman and Bonferroni-correctedWilcoxon tests with Cliff's delta effect sizes and bootstrap confidenceintervals; SHAP and permutation-importance analyses; and noise-sensitivity,per-insulator-type, residual, learning-curve and ablation studies. Running the script reproduces every quantitative result, figure and tablereported in the article and its Supplementary Information. The README mapseach generated output file to the corresponding figure or table. The training data are not redistributed here. The dataset is openly availablein Mendeley Data at https://doi.org/10.17632/8r7k4cgkg8.1 (sheet "Wholesamples", 6,656 records, four polymeric insulator types); download it intothe data/ directory before running. Requirements are pinned in requirements.txt. Python 3.10 or newer.Released under the MIT License.
Author: Burukanli, Mehmet
Institution: Bitlis Eren University
Language: English
Subfield: Electrical and Electronic Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7204803704