In this research, we investigated common machine learning algorithms to estimate the highest sensitivity of a D-shaped PCF SPR sensor by optimizing the performance parameters. Extreme gradient boosting (XGBoost), random forest model, and PyTorch neural network machine learning algorithms were compared to build a model and accurately predict the results, and sensor parameters were optimized through Non-dominated Sorting Genetic Algorithm (NSGA-II). The XGBoost technique demonstrated exceptional prediction capability, achieving an impressive R2 value of 99.64% and the trained model served as the objective function. The maximum sensitivity of 4529.75 nm/RIU was achieved in the standard optimization approach, However, with the guidance of NSGA-II, this sensitivity increased to 4814.14 nm/RIU, representing an improvement of 6.28%. The developed model enables rapid, reliable, and computationally cost-effective parameter predictions. Additionally, it provides a comprehensive understanding of the intricate relationships between input parameters and sensitivity, thus contributing significantly to the existing literature in the quest for optimal parameter identification through the application of machine learning algorithms.
Read full abstract