A novel data augmentation method and a data-driven prediction model for surface flashover at gas–solid interfaces under nanosecond pulses

Abstract Surface flashover at gas–solid interfaces is a critical factor in electromagnetic pulse simulator reliability. To accurately predict flashover events over a wide surface distance range (15–500 mm), this paper develops a machine-learning-based classification model. Three experimental platforms with output ranging from ± 80 kV to ± 2000 kV were constructed, yielding 1245 valid data samples covering various surface distances, voltage polarities, gas pressures, electrode configurations, and voltage waveforms. To address severe class imbalance (flashover proportion > 85%) under certain experimental conditions, a data augmentation method based on the three-parameter Weibull distribution is proposed: non-flashover samples are generated by sampling below a low cumulative probability threshold ( U 10% recommended) after estimating Weibull parameters of flashover voltages, effectively mitigating imbalance and overfitting. Six algorithms including Support Vector Machine (SVM), Multilayer Perceptron (MLP), Random Forest (RF), Gradient Boosting (GB), Extreme Gradient Boosting (XGBoost) and Light Gradient Boosting Machine (LightGBM) are trained with Bayesian hyperparameter optimization. The SVM model achieves the best performance on the test set: F1 score of 0.9111 and AUC of 0.9590. MLP achieves the second-best performance. The tree‑based ensemble methods show slightly lower F1-scores and exhibit a tendency towards overfitting. Feature importance and SHAP analysis are carried out to verify whether the model captures physically consistent mechanisms rather than spurious statistical correlations.