Machine learning flow regime classification in three-dimensional printed tubes

This paper presents how the machine learning (ML) approach combined with a three-dimensional (3D) printing technique facilitates the flow analysis in fluidic devices. A digital light processing 3D printing rapidly prototypes geometrically complex flow devices with low cost. Then a simple but powerful machine learning algorithm, the random forests algorithm, is used to classify the flow images taken from semitransparent 3D printed tubes. In particular, this work focuses on the laminar-turbulent transition process occurring in a 3D wavy tube and a straight tube, which is visualized by dye injection. The ML model automatically classifies experimentally obtained flow images within $\sim 0.01$ second per image only and identifies when and where the flow regime changes with an accuracy greater than 0.95. This work demonstrates a high-throughput and accurate method of flow visualization analysis.

Figure 1

Images of (a) a 3D printed straight tube, (b) a wavy tube, and (c) a DLP printer.

Figure 2

Flow images of a straight tube classified with flow regime labels: (a) label 0 (laminar flow only), (b) label 1 (laminar and transitional flow), (c) label 2 (laminar, transitional, and turbulent flow), (d) label 3 (transitional and turbulent flow), and (e) label 4 (turbulent flow only).

Figure 3

Flow images of a wavy tube classified with flow regime labels: (a) label 0 (laminar flow only), (b) label 2 (laminar, transitional, and turbulent flow), (c) label 3 (transitional and turbulent flow), and (d) label 4 (turbulent flow only).

Figure 4

Flow images of a wavy tube classified with the labels for transition onset location: (a) label 0 (laminar flow only), (b) label 1 (transition starts from the first section), (c) label 3 (transition starts from the third section), and (d) label 4 (transition starts from the fourth section).

Figure 5

Hyperparameter tuning of classifier 4 on (a) training and (b) validation data sets.

Figure 6

Flow regime variation with time in (a) a straight tube (classifier 1) and (b) a wavy tube (classifier 2). (c) Transition onset location variation with time in a wavy tube (classifier 4).