Accelerating photonic gas sensor design: machine learning-driven inverse optimization of silicon photonics Bragg gratings

Optical gas sensors based on photonic structures offer label-free, real-time detection with high sensitivity, but their design optimization remains computationally expensive. We present a hybrid machine learning framework that automates the inverse design of polymer-filled slot Bragg grating (PSBG) sensors for rapid prototyping across multiple target gases. Our approach combines Optuna-tuned artificial neural networks with a multi-model stacking ensemble (gradient boosting, random forest, XGBoost, ridge regression) to predict optimal structural parameters (grating period, depth, ridge width, slot height) from desired spectral characteristics. The framework incorporates domain-informed preprocessing (polynomial feature expansion, physics-based gas encoding) and per-target weighted meta-learners to ensure design fidelity. Validated on 1,045 design samples for CO2 and CH4 detection, the model achieves R2 > 0.99 for all parameters, enabling inverse design without iterative electromagnetic simulations. This work demonstrates the potential of ensemble learning for accelerating photonic device development and supports scalable platform-based sensor design.