Unsupervised Data Analysis of Direct Numerical Simulation of a Turbulent Flame via Local Principal Component Analysis and Procustes Analysis

Direct Numerical Simulations (DNS) of reacting flows provide high-fidelity data for combustion model reduction and validation, although their interpretation is not always straightforward because of the massive amount of information and the data high-dimensionality. In this work, a completely unsupervised algorithm for data analysis is investigated on a data-set obtained from a temporally-evolving DNS simulation of a reacting n-heptane jet in air. The proposed algorithm combines the Local Principal Component Analysis (LPCA) clustering algorithm with a variables selection algorithm via dimensionality reduction and Procustes Analysis. Unlike other data-analysis algorithms, it requires null or limited user expertise as all of its steps are unsupervised and solely entrusted to mathematical objective functions, without any hyperparameter tuning step required.