In the context of process and product design, predictive models are increasingly employed. Decomposition properties of chemicals may be experimentally determined through calorimetric measurements, and a few molecular structure-based models- which correlate the molecular structure of compounds with their decomposition properties- are also available. The aim of this paper is to improve predictive modeling of decomposition characteristics derived from Differential Scanning Calorimetry (DSC) (Baati et al., 2016), through the implementation of pattern recognition as a primary classification. For this purpose, the entire decomposition peaks of the molecules are represented and treated with image processing algorithms to identify the different patterns. Predictive modeling is then performed within the categories and compared to a global model prediction. Firstly, DSC thermograms (or curves) are analyzed to identify similar decomposition patterns in order to develop a clustering based on their overall thermal behavior instead of their structural similarities. Secondly, the reparation of the structural groups in the clusters is evaluated in order to determine the most influential functional groups on the thermal decomposition behavior. From this analysis, a systematic classification is developed to assign molecules of unknown thermal behavior to a particular cluster. Thirdly, predictive models of thermal characteristics are constructed within the different classes allowing predicting the entire DSC curve. The primary classification based on the pattern recognition increases the predictive performance of the regressions models (C) 2017 Institution of Chemical Engineers. Published by Elsevier B.V. All rights reserved.