Machine learning approach for accelerated reliability testing of integrated photonics

This study presents a machine learning-based approach to predict the reliability of Indium Phosphide (InP) Fabry–Pérot (F–P) lasers using accelerated life testing (ALT) data. InGaAsP F–P lasers with varying gain medium lengths (300, 500, and 750 μ m) from two different wafers were tested under thermal stress at 85 °C and different current densities. Light–Current–Voltage (LIV) measurements were performed after multiple stress periods (0, 24, 48, 72, 168, and 520/600 h) to monitor the threshold current ( I th ), which served as the primary indicator of device degradation. The model predicts the evolution of I th under combined thermal and electrical stress, enabling early identification of device failure without requiring long-duration tests. Several machine learning models, tree-based (Gradient Boosting (GB), Random Forest (RF), XGBoost), neural network (MLP), and linear regression (LR), were evaluated. The GB achieved the best predictive performance. The maximum deviation between predicted and measured I th remained around 15% across all stress periods, confirming that the model accurately captures I th evolution under cumulative stress. The proposed model is specifically optimized for the studied InP F–P laser dataset, representing the best fit for the available data and the most reliable framework for this device type. The proposed approach enables reliable prediction of laser failure after only 48 h, significantly reducing testing duration. This work demonstrates that machine learning offers a fast, reliable, and interpretable framework for predicting F–P laser lifetime.