Machine Learning for Predicting the Surface Plasmon Resonance of Perfect and Concave Gold Nanocubes

Using the combination of the discrete dipole approximation (DDA) and machine learning methods, we have developed a computational tool to predict the wavelength at which the dipole surface plasmon resonance (SPR) of gold concave nanocubes (GCNCs) takes place. First, we have used the DDA to generate SPR data considering two main features, the length and the concavity of the nanocube. Then, for training, test, and validation, two mechanisms were considered. Mechanism A consisted in splitting 100% of the generated data into two separate sets, one covering 75% and other with the remaining 25% of the whole data. The two separate subsets were used as training and test sets, respectively. Mechanism B basically consisted of SPR data set splitting into k subsets, following a k-fold cross validation procedure. For the machine learning algorithms, we used the K-nearest neighbors model, the ridge regression, and the artificial neural network regressor. The three models utilized here lead to different accuracies that depend on the selection of the mechanism used, either A or B. It was found that the accuracy in the prediction is sensitive to the L-2 regularization when the ridge regression is used. On the other hand, when the K-nearest neighbors model is employed, the accuracy depends on the number of nearest neighbors utilized during the calculations. With ridge regression or K-nearest neighbors, the accuracy obtained is between 80 and 94%. The best method for the SPR prediction of the GCNCs is the artificial neural network with eight neurons and L-2 = 10 because the accuracy for the predicted values is around 93%.