Vision algorithms are used not only in the computer vision domain but also in multifarious nuclear applications, e.g., flow visualization, medical image processing, multiphase analysis. We aim to develop in-house algorithms that establish a seamless connection between experiment and computation, leveraging image processing, computer vision, and other advanced AI techniques.
EXAMPLE 1: Automated boiling bubble analysis using CNN & optical flow
Fig. 1. Schematic of the automated boiling bubble analysis algorithms
A data-driven post-processing approach (Fig. 1) was proposed to segment, track, and identify wall-attached vapor bubbles from experimental high-speed video (HSV) images of subcooled flow boiling. Firstly, a transfer learning framework with a U-Net-based convolutional neural network (CNN) architecture was employed to detect and segment bubbles in HSV images with diverse contrast and surface texture. The segmented bubbles are further classified into individual and coalesced patches, with only the individual one tracked until it either condenses or coalesces with others. Then, a criterion to identify/exclude a condensing bubble (no longer contributing to heat transfer) based on the divergence of the bubble displacement was suggested; the displacement is calculated from sequential segmented bubble images using a global optical flow code. Using this combination of machine learning and optical flow, we can identify the full bubble life-cycle stages, i.e., nucleation, growth, sliding, and condensation.
Fig. 2. Nucleation activity map
Decrypting the bubble dynamics is crucial to understand boiling heat transfer mechanisms. Aforementioned algorithms facilitate obtaining related fundamental boiling parameters (e.g., nucleation site density, departure frequency, growth time, bubble departure diameter) which also can be input to computational model (see MULTIPHASE HEAT TRANSFER). For instance, the location of nucleation sites is identified by detecting circular object on the activity map obtained by time-averaging nucleation bubble patches (Fig. 2).
J. H. Seong, M. Ravichandran, G. Su, B. Phillips, M. Bucci*
Automated bubble analysis of high-speed subcooled flow boiling images using U-Net Transfer learning and global optical flow
International Journal of Multiphase Flow (IJMF), 159, 104336
EXAMPLE 2: (Deep learning based) optical flow algorithms to refine PIV velocity
Fig. 3. Instantaneous velocity fields of turbulent jet flow obtained by conventional PIV
cross-correlation method (left) and in-house global optical flow algorithm (right).
[PHASE 1] A global optical flow algorithm (a.k.a O-flow) was developed using MATLAB to enhance the velocity resolution of particle image velocimetry (PIV), allowing successful derivation of high-resolution fluid mechanical structures and quantities while preserving the conventional PIV results, as exemplified in the turbulent jet flow (Fig. 3).
[PHASE 2] The state-of-the-art CNN-based optical flow model was incorporated into the correlation-based coarse-to-fine approach. The CNN architecture was modified (Fig. 2) considering the boundary characteristics of PIV particle image pairs to produce more accurate results in high-gradient small-scale flow. A max pooling layer was used to prevent over-fitting. Then, the modified CNN model was trained in synthetic particle image datasets, generated with polynomial velocity profile which represent various types of local flows and real PIV particle conditions.
Fig. 2. Modification on CNN architecture
Fig. 3. Energy spectrum of turbulent jet fow
An energy spectrum of turbulent jet flow velocity was evaluated (Fig. 3) to quantitatively compare the performance of different methods. The spectrum obtained by the conventional correlation method collapses earlier in the small-scale region (i.e., high frequency) due to low-spatial resolution. In contrast, the spectra of the proposed method, i.e., CNN-based optical flow, and O-flow retain the inertia range. Note that, O-flow requires parameter (e.g., gamma) tuning to cope with variations in image brightness and produce optimal output. Meanwhile, the CNN-based algorithm does not require selecting optimal parameters. This has advantages over the classic method in reducing uncertainties associated with user-dependent processes. Those calculated velocity fields can be compared with CFD simulation.
J. H. Seong, M. S. Song, D. Nunez, A. Manera, E. S. Kim*
Velocity refienment of PIV using global optical flow
EIF, 60, 174
J. S. Choi, E. S. Kim*, J. H. Seong*
Deep learning based spatial refinement method for robust high-resolution PIV analysis
Experiments in Fluids (EIF], 64, 45
34141 대전광역시 유성구 대학로 291 (구성동 한국과학기술원 373-1) 기계공학동 www.kaist.ac.kr