Surrogate model for CFD based on machine learning

Abstract The paper presents the results of a data-driven surrogate modelling effort to reproduce the results of a Computational Fluid Dynamics (CFD) model. The CFD model considered here simulates a scenario where air containing CO 2 flows through a tube. A metal scaffold (that are highly selective in adsorbing Co2 from air) is placed inside the tube that absorbs CO 2 from the fluid as it flows through. The CFD model computes two important quantities: (i) the amount of transportation (a measure of CO 2 absorbed by the metal scaffold) and (ii) the amount of fluid mixing. Given the shape features of the metal scaffold as input, the CFD model (a) produces a 3D lattice filling part of it with the metal scaffold, (b) uses Lattice-Boltzmann equation to solve the CFD, and (c) compute amount of transportation and fluid mixing. The CFD model is time consuming. To reduce the computation time, we have investigated surrogate models for this CFD based on a machine learning (ML) method. The ML model takes as input the shape features of the metal scaffold and predicts amount of CO 2 absorption and fluid mixing. The objective is to come up with a ML model that produces the above quantities with reasonably low error. We investigated several ML models to find their effectiveness to produce reasonably accurate results for predicting the two quantities above. The data for training and testing the ML model was generated from the CFD based on different shape features of the metal scaffold. The underlying objective of the project was to produce an optimal design of the metal scaffold that maximises absorption (of CO 2 ) and fluid mixing. The optimization was done using an evolutionary algorithm (EA). EA starts with a set (population) of random shapes and modifies the shapes over generations based on feedback from CFD. In each generation the CFD computes the absorption and mixing for each of the shapes in the population. The best of the shapes from the population goes through some transformations (crossover and mutation) to generate the population for next generation. The process continues until the optimization converges. The shape features and their fitness (absorption and mixing) of the population from the several generations are combined into one large dataset where the shape features are input, and the fitness values are target. This constitutes the data used for training and testing the ML models. Given a total of n generations, we trained ML models on data up to k –th generation and tested the model on the data from last n–k generations. We present the CFD model, the surrogate models, and some preliminary results in this paper.