A Comparison of Optimizers in a PyTorch Based Artificial Neural Network to Predict Normal Boiling Points of Alkanes

In this study, we explore the effectiveness of various optimization algorithms for predicting the normal boiling points of alkanes using PyTorch-based neural network models. Alkane molecular structures were uniquely encoded into a machinereadable format and used to train both an artificial neural network (ANN) and a graph neural network (GNN) model. We compared the performance of three optimizers; Stochastic Gradient Descent (SGD), Adam, and Rprop, across two loss functions (L1Loss and MSELoss) and multiple learning rates (0.01, 0.001, and 0.0001). Our results indicate that Adam consistently yields the lowest mean squared error (MSE) values across both models, highlighting its robust performance. Notably, in the ANN framework, Rprop demonstrated rapid convergence and outperformed SGD, whereas in the GNN model, SGD showed superior performance compared to Rprop.The analysis of MSE values further corroborated these findings, emphasizing the critical role of optimizer selection in achieving both high accuracy and model stability. These insights provide valuable guidance for future applications of neural networks in chemical property prediction and underscore the importance of fine-tuning optimization parameters in machine learning workflows.