Machine learning techniques for predicting the optical properties of a biosensor

This study utilizes machine learning algorithms to accurately predict the optical properties of a biosensor. To enhance the sensitivity of the sensor, a gold layer and an MXene coating are incorporated as crucial components of the design to excite the surface plasmon resonance (SPR) phenomena. The simulations show that the sensor effectively detects cancer cells, blood samples, and viral biomarkers such as DNA and RNA, covering refractive indices ranging from 1.30 to 1.43. Since no experimental or publicly available scientific dataset was accessible, a total of 1056 data samples were generated through COMSOL Multiphysics simulations by systematically varying the analyte refractive index, operating wavelength, gold layer thickness, and MXene layer thickness. Our models focus on four key optical properties—sensitivity, effective refractive index N eff , effective mode area A eff , and confinement loss. Different machine learning models, including regression models and neural networks, were trained and evaluated using 10-fold cross-validation, with performance assessed in terms of mean squared error (MSE) and the coefficient of determination R 2 . A comparison of the artificial neural network with other methods reveals that the performance of the ANN is superior for both the full and half datasets. Our findings indicate that the ANN produces predictions in strong agreement with the sensor sensitivity test data. It achieves the best performance with a test MSE of 9.8×10 −5 and a test R 2 score of 0.999907. The decision tree regressor (DTR) also predicts well with a test MSE of 0.008813 and a test R 2 score of 0.991603. Almost the same trend is observed for the effective refractive index, effective mode area, and confinement loss. This work demonstrates that machine learning significantly reduces simulation time compared to COMSOL and can serve as an effective alternative to manual design optimization.