Identification of flow regimes in boiling flow with clustering algorithms: An interpretable machine-learning perspective

Longxiang Zhu, Zhiee Jhia Ooi, Taiyang Zhang, Caleb S. Brooks, Liangming Pan

Research output: Contribution to journalArticlepeer-review

Abstract

The flow regime is the prerequisite to accurately modeling two-phase flow. Unsupervised machine learning techniques enable the identification of flow regimes objectively. Previous machine learning models are used as a “black box” tool without knowing the physical phenomena in the flow regime. Consequently, the cause of the identification error tends to be poorly understood and the model cannot be fundamentally improved. The paper develops an approach to better understand the identification result by creating a mapping relation between bubble distribution and the components in machine learning algorithms. The intrinsic interpretation generates the clustering principle to guide the feed-in feature extraction and clustering algorithm selection processes. Four features extracted from the bubble-size raw data recorded using conductivity probes are examined. Among them, the Cumulative Distribution Function of the chord length in seven dimensions is demonstrated to be the appropriate feed-in feature. Three major kinds of clustering algorithms are investigated, including partition-based, hierarchy-based, and model-based methods. After assigning physical meanings to the nodes in the algorithm and inspecting the clustering outcomes, the K-means, K-medoids, and Self-Organizing Maps are shown to succeed in the flow-regime identification problem. In addition, the local and the global flow regimes are generated by the well-designed machine learning model to assist the understanding of the boiling flow structure in a multi-dimensional way and in an area-averaged sense. The overall accuracy of the machine learning model for the three global flow regimes is 86%, which suggests the chosen algorithm with the selected feed-in feature is capable to capture the flow regime in the boiling flow. The flow regime map for the boiling dataset is compared with the existing flow regime criteria developed in the air–water flow, the result of which highlights the necessity of a new criterion to capture the transition from bubbly to slug for boiling flow. For the range of flow conditions considered in this work, the transition criterion between bubbly and slug flows is proposed to be 0.14 for upward boiling flow in an annular channel.

Original languageEnglish (US)
Article number120493
JournalApplied Thermal Engineering
Volume228
DOIs
StatePublished - Jun 25 2023

Keywords

  • Boiling flow
  • Clustering algorithm
  • Flow regime
  • Interpretable machine learning
  • Two-phase flow

ASJC Scopus subject areas

  • Mechanical Engineering
  • Energy Engineering and Power Technology
  • Fluid Flow and Transfer Processes
  • Industrial and Manufacturing Engineering

Fingerprint

Dive into the research topics of 'Identification of flow regimes in boiling flow with clustering algorithms: An interpretable machine-learning perspective'. Together they form a unique fingerprint.

Cite this