Research on prediction and classification of physical and chemical properties based on neural network and regression model

This paper explores the optimization of prediction and classification methods for physical and chemical properties in the Robot Chemist platform, employing polynomial fitting, BP neural networks, LSTM models, and logistic regression. First, missing data are filled using averages, and outliers are identified and removed through quartile analysis. A cubic polynomial model is then applied to predict the nonlinear relationship between molecule ID and the property y2. For the prediction of y1, a BP neural network with two hidden layers is used, achieving an accuracy of R<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> = 1.0. The same model structure is adapted to predict y3, with x1 and x3 identified as key influencing factors through correlation analysis. Logistic regression is employed for classification based on y1-y3 and x1-x100, achieving a classification accuracy of 0.897 and highlighting six main influencing features. Further optimization using LSTM models significantly improves the prediction accuracy for y1 and y3, reducing loss values, while reclassification using BP neural networks enhances classification accuracy to 0.944. These results demonstrate the effectiveness of deep learning models in improving the performance of predictive and classification tasks within the Robot Chemist platform.