Desenvolvimento de redes neurais artificiais para resolução de problemas inversos em cinética química: termogravimetria não isotérmica e motores de combustão interna

Abstract

Inverse problems are common in chemical kinetics and often lack unique or stable solutions, especially when associated with experimental noise and multistep reaction mechanisms. Traditional kinetic model-fitting methods face significant limitations under such complexity. In this context, this thesis proposes the development of a MultiLayer Perceptron (MLP) artificial neural network to solve inverse problems in thermal analysis, focusing on the determination of the kinetic triplet — activation energy, pre-exponential factor, and reaction model — from non-isothermal thermogravimetric data and internal combustion engine experiments. The methodology was validated using simulated data and applied to real systems of high scientific and technological relevance, including biopolymers, antioxidants, nanomaterials, antiviral drugs, and automotive fuels. In the context of thermal analysis, the MLP network effectively combined multiple kinetic models, providing a more realistic interpretation of decomposition processes. For internal combustion engines, the network was adapted to analyze burned mass fraction (MBF) data, allowing the inference of combustion kinetic parameters under real operating conditions. The results demonstrate that the proposed MLP network is a robust, accurate, and versatile tool for kinetic modeling in complex systems. The hybrid approach preserves the physical significance of the parameters and expands the potential of artificial intelligence applications in chemistry and engineering, offering a new paradigm for kinetic studies in both academic and industrial environments.