Research On LSPR Sensing Performance Based On Machine Learnin

When light of a specific wavelength is incident on the surface of a metal nanostructure,it can excite the phenomenon of localized surface plasmon resonance（LSPR）,and the position of the absorption peak will shift with the change of the refractive index of the dielectric on the surface of the metal nanostructure.Therefore,metal nanostructures are widely used as refractive index biosensors.The performance of LSPR biosensors based on refractive index is related to the material,shape,size,and arrangement of the basic units of the metal nanostructure.The design and fabrication process of traditional LSPR biosensors relies on large-scale simulation calculations,which are time-consuming,computationally expensive,and inefficient.Furthermore,optimizing the sensitivity and reverse design work rely heavily on the designer’s prior experience and have issues such as non-unique solutions.In recent years,with the development of the field of artificial intelligence,using machine learning networks instead of traditional numerical simulation methods for optical structure design can significantly reduce time costs.However,most of these design work focus on learning the mapping relationship between metal nanostructures and spectra.In the field of LSPR refractive index biosensor design,there is still a lack of machine learning-based methods for forward prediction of LSPR sensing performance and reverse design of nanostructures.To address these issues,this study mainly includes the following three aspects:（1）A machine learning-based method for predicting the sensing performance of metal nanostructures is proposed.In this study,a dataset of gold nanostructures and sensing performance is established for machine learning model training.Five machine learning networks were used to predict the refractive index sensitivity,full width at half maximum,and quality factor of different gold nanostructures.Among them,the random forest and backpropagation neural network performed better,with prediction errors of less than 10nm/RIU,6nm,and 0.8RIU^-1 for refractive index sensitivity,full width at half maximum,and quality factor,respectively.The prediction efficiency of random forest is 60 times higher than that of traditional simulation methods,proving the effectiveness of the machine learning-based forward prediction method for LSPR sensing performance.（2）A reverse design method for LSPR sensors combining backpropagation neural network with genetic algorithm and multi-objective particle swarm algorithm is proposed.By defining the fitness function of refractive index sensitivity,full width at half maximum,and quality factor,the ideal gold nanostructure is designed,and the effectiveness of the method is verified using the finite-difference time-domain numerical simulation method.The proposed design method solves the problems of traditional reverse design process that relies on prior knowledge,high time cost,and non-unique solutions.（3）A method combining multi-objective particle swarm algorithm with backpropagation neural network is used to complete the reverse design of the gold nanorod LSPR sensing structure by selecting refractive index sensitivity,full width at half maximum,and quality factor as optimization targets.A triangular arrangement array of gold nanorods with excellent sensing performance was designed within 2 minutes,with evaluation indicators of 486.7nm/RIU,53.02nm,and 9.18 RIU^-1.Compared with traditional reverse design methods,80%of the computation time is saved.