Artificial neural networks for predicting the viscosity of lead-containing glasses
This work addresses a specific challenge in materials science for glass manufacturing, but it is incremental as it applies an existing method (neural networks) to a new dataset with improved performance over prior models.
The researchers tackled the problem of predicting the viscosity of lead-containing glasses, which is crucial for manufacturing, by using artificial neural networks trained on the SciGlass database, resulting in a model that outperformed seven existing models in terms of mean absolute error and coefficient of determination on test data.
The viscosity of lead-containing glasses is of fundamental importance for the manufacturing process, and can be predicted by algorithms such as artificial neural networks. The SciGlass database was used to provide training, validation and test data of chemical composition, temperature and viscosity for the construction of artificial neural networks with node variation in the hidden layer. The best model built with training data and validation data was compared with 7 other models from the literature, demonstrating better statistical evaluations of mean absolute error and coefficient of determination to the test data, with subsequent sensitivity analysis in agreement with the literature. Skewness and kurtosis were calculated and there is a good correlation between the values predicted by the best neural network built with the test data.