Abstract
This research aims to enhance the deep neural network (DNN) model previously developed by this group, as referenced in(10), by integrating degradation time into its feature set. It specifically addresses the thermal decomposition of polyvinyl alcohol (PVA) at low heating rates of 2, 5, and 10 °C.min(-1). In addition, the study presents a thermo-kinetic analysis of the data, facilitating the estimation of activation energy and activation enthalpy. The inputs to the DNN frameworks include degradation time, degradation temperature, and heating rate. Modifications were made to the DNN model to tackle overfitting and reduce the discrepancy between output signals and experimental scatter. This was accomplished through iterative adjustments to the learning rate, implementation of data augmentation techniques, prolongation of the training duration, and early termination to minimize error. An optimized DNN architecture, comprising two hidden layers and eight neurons, effectively facilitated the learning algorithms and successfully trained arbitrary constants. This resulted in output signals that closely aligned with the experimental data ([Formula: see text]), thereby providing a ranking of parameter sensitivity characterized by heating rate, time, and degradation temperature. The Flynn-Wall-Ozawa (FWO) and Kissinger-Akahira-Sunose (KAS) model-free equations were used to estimate the activation energy ([Formula: see text]) of the thermogravimetric (TGA) data curves. The estimated average [Formula: see text] values derived from the FWO and KAS model-free equations were 64.6±3.2 kJ·mol⁻¹ and 58.8±2.9 kJ·mol⁻¹, respectively, based on conversion rates between 5 and 50 wt% (i.e., where [Formula: see text]). The estimated theoretical value of the activation enthalpy (ΔH) required for the formation of the activation complex, at these higher correlations, was determined to be positive (25.4- 102.0 kJ·mol⁻¹), indicating that the reaction is invariably endothermic and is consistent with established information in the literature. This approach could prove pivotal for manufacturers in the design and fabrication of polyvinyl alcohol (PVA) and composites with enhanced and novel properties.