Supervised machine learning for identification of glass properties: Towards structural Stability and performance

Abstract

Abstract Classical and physics-based modelling is a basic way to describe how physical processes work, but it has many problems. For example, it uses a lot of computing power, takes a long time, and can't show how random and complicated processes work in glass science and engineering. On the other hand, machine learning (ML) models have been shown to get around this problem, especially when a precise and reliable estimate is needed. In this study, neural network (NN), adaptive neuro fuzzy inference system (ANFIS), k-nearest neighbors (KNN), and robust linear regression (RLR) models were used to simulate the spring constant (K) at the junction of structural glass plates. The data from the experiment, which included axial load (N) and four different displacements (mm) and was collected in a total of 2879 cases, was pre-processed and split into 70% calibration and 30% verification. After that, sensitivity analysis was done, and 6 different model combinations (M1 through M6) were made. Based on the results of three performance evaluation criteria (R2, RMSE, and R), the ML model did well and could be trusted to estimate K. The ANN-M5, ANN-M6, ANFIS-M5, ANFIS-M6, KNN-M5, KNN-M6, RLR-M5, and RLR-M6 models, on the other hand, did 0.1 percent better than the rest. The model follows the latest best practices in machine learning and makes it possible to do experiments on low-power edge computing devices with minimal cost. KNN-M5 and KNN-M6 were the best models in terms of RMSE, but the confidence interval values showed that they were better than the best model (95%).