Unsupervised Learning in Turbulent Combustion

Background:

Machine learning (ML) and Artificial Intelligence (AI) are new-age tools that have changed the way we analyze, interpret, and use data in almost all fields - scientific and non-scientific. In scientific domains too the use of ML has drastically increased, and it is giving significant insights and new modeling paradigms that are shaping the way for new-age data interpretation and data-driven modeling. In this research theme, we apply available ML tools (supervised as well as unsupervised) to understand and model flow problems (reacting as well as non-reacting). We confirm domain knowledge that is already known from decades of research or identify new insights that enhances our understanding. We modify the ML algorithms to include physical observations or laws into the ML training procedure such that we make balanced use of available data and available laws to be obeyed.

Why did we solve this problem?

Combustion happens in several different ways - conventional combustion, Moderate & Intense Low oxygen Dilution (MILD), high-temperature combustion etc. In MILD combustion, which is also known as Flameless Combustion, there is no visible flame. Analyzing such flames and identifying scalars (chemical markers) that best correlate with the heat release rate (HRR) is essential to understanding flame location, topology, etc. Similarly in other ways of combustion identifying different regions of reacting flow fields is essential not only to understand but also to develop local modeling strategy that are accurate with low computational cost. Thus, it is essential to interpret reacting flow fields, diagnose them, and characterize them to understand several ways in which combustion happens, troubleshoot, and develop reduced-order models.

What did we find?

When thermochemical data is fed into clustering-cum-dimensionality reduction algorithms such as Vector Quantization using Principal Component Analysis (VQPCA) we are able to identify important clusters inside reacting fields that correlate with the intensity of heat release rate (HRR). Specifically, in MILD combustion we found that when oxygen-dilution levels are low, i.e., there is relatively still enough oxygen, say 3.5% by volume, reaction zones are still relatively strong and the variation of reaction scalars within heat releasing regions is largely due to chemical reactions. However, when MILD combustion fields have higher oxygen dilution levels, say 2.0% by volume, reaction zones weaken and variation in reaction scalars can be due to both chemical reactions and diffusion phenomena. In such cases, ML algorithms like Vector Quantization using Principal Component Analysis (VQPCA) reach limitations in detecting reacting structures unambiguously and we need to be careful as to what data we feed into VQPCA. We developed a physics-based approach to identify chemical markers that help selecting the variable that we feed. In a subsequent article, Savarese et al. developed a new index that helps in deciding how many clusters are appropriate to partition reacting flow fields such that we balance interpretation with accuracy of reduced-order models.

Impact

The impact of these studies is that it helps in improving our understanding of complex reacting flow fields helping use to distill key chemical markers that correlate best with heat release rates. Secondly, they help in identifying the most important and appropriate number of regions in reacting flows that can help improve local modeling strategies in numerical combustion modeling.

You can read more about this topic in the following references. Additionally, if this research theme piques your interest and you want to know more, learn more, and contribute more on this topic, please feel free to write.

References:

[1] Dave, H., Swaminathan, N., & Parente, A. (2022). Interpretation and characterization of MILD combustion data using unsupervised clustering informed by physics-based, domain expertise. Combustion and Flame, 240, 111954.

[2] Savarese, M., Jung, K. S., Dave, H., Chen, J. H., & Parente, A. (2024). A new index for the comparative evaluation of combustion local low-dimensional manifolds. Combustion and Flame, 265, 113434.