Due to air pollution, CO2 is commonly present in the air. Substrates can be designed to absorb CO2 to make air free from pollutant. Fluid dynamics model based computer simulations are commonly used to design substrates and understand the amount of absorption the substrate can do. In this paper, we explored the use of deep learning model as an efficient surrogate for Computational Fluid Dynamics (CFD) simulations. CFDs are often constrained by high computational costs. Our investigation centers on a specific scenario involving the flow of air containing CO2 through a tube with a metal scaffold (i.e., substrate) designed for selective CO2 adsorption. We developed a deep learning model capable of accurately predicting transport (i.e., absorption) and mixing properties from the geometric characteristics of the substrate, thereby circumventing the need for computationally intensive CFD simulations. Our model employs a 3 Dimensional (3D) Convolutional Neural Network (CNN) to process the discretized 3D shapes of the tube and scaffold. We validated this approach using data generated across multiple runs of CFD as part of an optimization process. Our findings demonstrate that our deep learning models can predict transport and mixing properties with high accuracy, significantly reducing computation time compared to traditional CFD methods.