Unsupervised Clustering: A Mixture of Experts Framework to Represent Flamelet Tables
Abstract
A novel unsupervised learning-based clustering approach to represent the flamelet tables is developed. The typical tabulation method for flamelet-based modeling generally requires a large amount of storage. A well-developed machine learning model can accurately represent flamelet tables while taking up significantly less storage. The proposed method utilizes a mixture of experts (MoE) technique where specialized Deep Neural Networks (DNNs) are trained on different parts of the input space. This identification of combustion manifolds within the input space is accomplished through the use of an unsupervised learning-based clustering algorithm, which is able to categorize an input to a specific cluster. Previous studies have shown that developing specialized models can lead to a higher accuracy and faster access to the flamelet tables. However, the clustering techniques utilized in these studies do not investigate an unsupervised learning approach. The proposed model is trained and evaluated on 5-dimensional flamelet tables, and an investigation of clustering techniques and optimal number of clusters is also conducted. This research project shows that unsupervised learning-based clustering algorithms coupled with a MoE framework of DNNs can accurately predict temperatures and mass fractions in flamelet tables.
Subject
FlameletFlamelet Table
Mixture of Experts
MoE
Machine Learning
Deep Learning
DNN
Neural Network
Clustering
Unsupervised Clustering
Combustion
Citation
Mayilvahanan, Sarvesh (2022). Unsupervised Clustering: A Mixture of Experts Framework to Represent Flamelet Tables. Undergraduate Research Scholars Program. Available electronically from https : / /hdl .handle .net /1969 .1 /194397.