PHYSICALLY INFORMED NEURAL NETWORKS TO REPRESENT MOTION OF GRANULAR MATERIAL

The growing importance of renewable energy sources, particularly biomass, in mitigating climate change has led to increased research and development in this field. Biomass combustion chambers play a crucial role in converting biomass into heat energy efficiently and cleanly. However, the combustion process is influenced by the characteristics of biomass particles, such as their composition, size distribution, and moisture content, which can vary significantly. In this thesis, we explore the application of physics-informed neural networks (PINNs) for predicting particle motion in biomass systems, aiming to improve combustion efficiency and reduce emissions. The thesis is divided into three main parts. The first part introduces a PINN model for predicting biomass particle velocities in a moving bed. The proposed model demonstrates the effectiveness of PINNs as surrogate models for computationally expensive discrete element method (DEM) simulations. The second part focuses on reconstructing biomass particle fields in a particle-fluid problem using continuum methods. Through the integration of simulation data and PINNs, accurate predictions of particle velocities, pressure, and density fields are achieved. In the third part, a recurrent neural network architecture is employed to predict particle motion in a moving grate chamber and a rotating drum. The results highlight the model’s ability to capture particle behavior accurately, even for time intervals longer than the training time. The findings of this thesis contribute to the understanding of biomass combustion dynamics and offer valuable insights into optimizing biomass combustion chamber design and operation. The PINN-based approach proves to be computationally efficient and reliable for predicting particle motion, providing a valuable tool for biomass system analysis and optimization. Additionally, a separate study on three-dimensional computational fluid dynamics - discrete element method (CFD-DEM) simulations of raceway transport phenomena in a blast furnace is presented, emphasizing the importance of understanding the complex dynamics in metallurgical processes. By combining the power of neural networks with physics-based modeling, this research bridges the gap between computational efficiency and accuracy in predicting particle motion in biomass systems. These insights can contribute to the development of cleaner and more efficient biomass combustion technologies, promoting a sustainable and greener future.