Dissertation Proposal - Asif Rahman

Physics-informed Neural Network for Predicting Chemo-Physical Phenomena of the Cement-based Materials

Presented by Asif Rahman – Data Science emphasis

Virtual – Click here to join via Zoom

Abstract

Portland cement concrete is the mostly used construction material on Earth. However, many structures built with cementitious materials face premature degradation problems. Cement-based materials are porous and hence, exposure to aggressive chemicals such as acids, sulfates, and chlorides etc. can cause detrimental chemical reactions in the presence of water and/or oxygen. This will lead to loss of structural integrity and subsequent damage with the progression of time. As such, it is crucial to capture the proper kinetics involved in the detrimental chemical reactions, to conduct the durability studies. Several researches have been conducted to run multi-scale chemo-physical simulations on these chemical attacks in the cement-based materials. Most of these simulations rely on the traditional finite element solvers, which require complex mesh generation and mathematical derivations, leading to high computational cost. Moreover, these simulations rely on specific model parameters which can raise more errors from a little change in the material properties, while extrapolating the data. Within this context, my Ph.D. proposal aims to develop a physics-informed deep learning framework to incorporate the chemo-physical behavior of the cement-based materials. To facilitate this, an original physics-informed neural network (PINN) architecture is proposed to enable a data-driven solution, to predict cement hydration temperature (CHT) and internal sulfate attack (ISA) in the cement-based materials. The proposed PINN-CHT and PINN-ISA will be able to capture the relevant chemical reactions that can be interpreted by the law of physics. These developments will guide the materials engineer to identify the chemical process as well as to predict the in-field performance of concrete and cement-based materials. There is potential opportunity to infer the knowledge and algorithms from these PINN models to predict new material properties and durability studies.

Committee

Dr. Yang Lu, Chair, Dr. Tim Andersen, Dr. Edoardo Serra