Purpose: To address the limitations of existing linear and exponential models in reflecting outliers and nonlinear characteristics inherent in transparent OLED degradation data, this study proposes a new degradation analysis method that combines an outlier removal technique with Symbolic Regression (SR).
Methods: Based on a full factorial design using UV, color, brightness, and UV blocking film as variables, accelerated life testing was performed under 32 conditions for 2,166 hours, securing a total of 4,992 relative luminance data points. After applying the RANSAC algorithm to the training data (0-2,030 h) to identify and remove outliers, linear, exponential, and SR degradation models were developed, and their performance was evaluated.
Results: The RANSAC algorithm effectively removed outliers, corresponding to approximately 7.5% of the total data. The subsequent SR modeling autonomously identified diverse physical degradation forms (e.g., linear, square root, and exponential-composite patterns) inherent in the data. By optimizing the complexity penalty through the Bayesian Information Criterion (BIC), the RANSAC-SR framework achieved a Pareto-optimal structure (Complexity 5.0). While maintaining comparable training errors to conventional linear models, it demonstrated superior extrapolation performance with a Mean Absolute Percentage Error (MAPE) of 5.85% at the 2,166 h unobserved validation point, thereby strongly avoiding overfitting to local noise. Notably, this performance was achieved using only 80% of the available training data, while baseline models utilized 100%, demonstrating superior data efficiency.
Conclusion: The proposed RANSAC-SR model overcomes the limitations of conventional outlier processing and modeling. This study presents a method to enhance prediction accuracy by reconstructing degradation over time into a physically interpretable continuous model, which is expected to be applicable to the reliability assessment of various displays and materials in the future.