Four-wave Mixing Spectroscopy Of Graphene And Its Heterojunctions Based On Machine Learning Classification Algorithms

With the continuous development of the semiconductor industry,the feature size miniaturization of chip devices has approached the bottleneck.In order to continue the"Moore’s Law"development model,both academic and industrial circles are constantly looking for breakthroughs in technology and new semiconductor materials.As a kind of atomically thick crystal,two-dimensional materials stand out among many semiconductor materials due to their excellent optoelectronic properties.In the process of applying two-dimensional materials to the preparation of optoelectronic chips,accurate characterization of their physical and chemical properties is an important task.Taking the number of layers of two-dimensional materials as an example,traditional characterization methods such as atomic force microscope and scanning electron microscopy and Raman spectroscopy provide us with a variety of options.But their shortcomings such as time-consuming and cumbersome measurement environment requirements,as well as damage to the sample and limited recognition range make the characterization of two-dimensional materials difficult.With the development of nonlinear optics,we realize that the nonlinear optical effects of two-dimensional materials are closely related to their nonlinear properties.By extracting the characteristic information of the nonlinear spectrum of two-dimensional materials,it can help us to quickly and accurately identify the number of layers and other information.In order to achieve this goal,we innovatively propose in this article to use machine learning classification algorithms to learn and train the model of the nonlinear spectral characteristic data set of two-dimensional materials,so as to identify unknown spectral information and accurately determine the physical and chemical properties of the corresponding two-dimensional materials.Taking graphene and its heterojunction samples as examples,this article mainly realizes the work of identifying and judging the number of layers.The main contents are:（1）Multiple graphene samples with 1～4 layers were prepared by mechanical peeling method,and the number of layers was calibrated by atomic force microscope and Raman spectroscopy;by dry transferring,the single-layer and two-layer graphene samples obtained by mechanical exfoliation were respectively prepared into two different heterojunctions with the five-layer and two-layer Mo S₂ film.The self-built nonlinear spectrum acquisition system is used to collect single-peak coherent anti-Stokes Raman scattering spectrum of elemental graphene samples and single-peak four-wave mixing spectrum of heterojunction samples.（2）Perform Gaussian fitting on graphene coherent anti-Stokes Raman scattering spectral data,extract center wavelength,peak area,half-height width and intensity data as features,and form a data set for model training and testing.Two machine learning models,support vector machine and gaussian process classification are used for training,and the test accuracy on the test set is 91.7%and 92.4%,respectively.Through the Gaussian process classification model of the anisotropic kernel,the characteristic weight of the spectral data is analyzed,and the main basis for the identification of the number of layers is obtained.A careful analysis of the reasons for its misjudgment revealed that the change and disturbance of the experimental environment（mainly the light spot shift caused by the micro-movement of the optical platform）caused the data overlap,which is the main reason for the misjudgment.It will be improved through fixed integration of the system optical path in the future.In addition,the existing optical system will be upgraded to improve the detection efficiency of coherent anti-Stokes Raman scattering spectrum in the Raman wavelet number region.（3）This method was verified on graphene/molybdenum sulfide heterojunction samples.The Gaussian fitting method was also used to fit the four-wave mixing spectrum data of the heterojunction,and the features were extracted.Support vector machine and gaussian process classification models are obtained through training,and the test accuracy obtained of both can reach 90%.Through the statistical average of the spectral data after Gaussian fitting,the characteristic changes of the two kinds of heterojunction four-wave mixing spectral data are analyzed.