AI-Based Video Analysis of Internal Two-Phase Flow in a Multiphase Pump

IFP Energies nouvelles (IFPEN) est un organisme public de recherche, d’innovation et de formation dont la mission est de développer des technologies performantes, économiques, propres et durables dans les domaines de l’énergie, du transport et de l’environnement. 

IFPEN met à disposition de ses chercheurs un environnement de recherche stimulant, avec des équipements de laboratoire et des moyens de calcul très performants.

Dans le cadre de la mission d’intérêt général confiée par les pouvoirs publics, IFPEN concentre ses efforts sur :

• l’apport de solutions aux défis sociétaux de l’énergie et du climat, en favorisant la transition vers une mobilité durable et l’émergence d’un mix énergétique plus diversifié ;
• la création de richesse et d’emplois, en soutenant l’activité économique française et européenne et la compétitivité des filières industrielles associées.

AI-Based Video Analysis of Internal Two-Phase Flow in a Multiphase Pump

Multiphase pumping plays a strategic role in industrial systems where gas–liquid mixtures must be transported efficiently without prior phase separation. Such configurations are increasingly found in emerging process applications where compactness, operational flexibility, and reduced process complexity are key advantages, such as Carbon Capture, Utilization, and Storage (CCUS) systems. However, despite their technological interest, the internal flow physics of multiphase pumps remain only partially understood, especially under high gas volume fractions where numerical approaches such as CFD become unreliable or difficult to validate.

To address this scientific and technological challenge, an instrumented multiphase pump test rig at IFPEN has optical access for high-frequency flow visualization. This unique setup creates a rare opportunity to investigate experimentally the internal two-phase flow patterns of a multiphase pump. At the same time, each experimental campaign generates a massive quantity of high-frequency videos, making manual analysis extremely time-consuming and preventing systematic, objective comparison between operating points. This motivates the development of advanced analysis methods capable of turning raw image data into exploitable physical information.

To go beyond qualitative observation and access the actual dynamics of the flow inside the pump, artificial intelligence offers promising perspectives for data-driven flow tracking and analysis. Recovering the motion of gas–liquid structures from high-frequency visualizations is, however, a difficult problem: the image content is dominated by deforming interfaces, strong intensity gradients and out-of-plane motion, conditions under which conventional correlation-based PIV loses robustness. Deep-learning optical-flow architectures have emerged as an attractive alternative in such regimes, as they estimate dense displacement fields without relying on local correlation assumptions.

The main obstacle to their use is the absence of ground truth: no reference velocity field is available for experimental recordings, and supervised training therefore requires synthetic image pairs with prescribed displacements. Generating such data from high-fidelity CFD is costly and, at high gas volume fractions, difficult to validate. Generative approaches such as Denoising Diffusion Probabilistic Models (DDPMs) open another route, by producing image content that reproduces the visual signature of experimental two phase flows at a fraction of the cost, and thus by helping to bridge the gap between synthetic training data and real acquisitions.

Work description 

To develop and evaluate a proof-of-concept video analysis pipeline for automatically characterizing internal two-phase flow structures in a multiphase pump from high-speed visualization data.

• Develop image and video processing methods to automatically detect, identify, and track bubbles and relevant flow structures in high-speed recordings of the pump.
• Extract quantitative information from the videos, such as flow organization, characteristic structure locations, and local velocity-related indicators using bubbles as natural tracers.
• Compare operating conditions objectively by defining metrics able to distinguish flow regimes and characterize the evolution of internal flow patterns.
• Investigate correlations between video-based observables, high-frequency pressure measurements, and global pump performance.
• Assess the ability of the proposed approach to support flow regime classification, improve understanding of flow/performance relationships, and provide useful indicators for pump design comparison.

Bibliography

• Gundersen T. Ø. S., Moënne-Loccoz V., Dupoiron M., Torbergsen E. A., Balakin B., Arntzen B. J., Hoffmann A. C. — Visualization of Unsteady Flow in a Multistage Helico-Axial Pump. Proceedings of the 15th European Conference on Turbomachinery Fluid Dynamics & Thermodynamics (ETC15), Paper ETC2023-168 (2023). DOI: 10.29008/ETC2023-168. [Experimental visualization of unsteady internal flow in a transparent three-stage helico-axial multiphase pump using high-speed imaging at 5000 fps; visualization of tip-leakage vortices, diffuser recirculation and secondary-flow structures, with comparison to URANS simulations.]
• Gundersen T. Ø. S., Dupoiron M., Moënne-Loccoz V., McAlinden T., Juliussen Y. C., Torbergsen E. A., Meredith A., Marthinussen S.-A., Arntzen B. J., Hoffmann A. C. — Commissioning A New Multiphase Pump Visualization Test Rig To Investigate The Internal Flow Field And Its Connection With Pump Performance. Proceedings of the 38th International Pump Users Symposium, Turbomachinery Laboratory, Texas A&M Engineering Experiment Station (2022). [Development and commissioning of a transparent multiphase-pump test rig dedicated to optical investigation of internal flow; establishes an experimental methodology for relating gas distribution, tip leakage and recirculation to multiphase pump performance.]
• Qianyu Zhu, Junjie Wang, Jeremiah Hu, Jia Ai, Yong Lee — PIV-FlowDiffuser:Transfer learning-based denoising diffusion models for PIV. Sensors 25, no. 19: 6077. DOI: 10.3390/s25196077 [PIV-FlowDiffuser is a transfer-learned denoising diffusion model (pretrained on Sintel/KITTI, fine-tuned on upsampled synthetic PIV) that removes residual noise in PIV vectors, cutting AEE by 59.4% vs RAFT256-PIV and improving generalization.
• Zhang, M., Piggott, M.D. — Unsupervised Learning of Particle Image Velocimetry. In: Jagode, H., Anzt, H., Juckeland, G., Ltaief, H. (eds) High Performance Computing. ISC High Performance 2020. Lecture Notes in Computer Science(), vol 12321. Springer, Cham. DOI:  10.1007/978-3-030 59851-8_7  [PIV-FlowDiffuser, pretrained on optical-flow datasets (Sintel/KITTI) and fine-tuned on upsampled synthetic PIV, uses a denoising diffusion process to remove residual noise in flow fields—cutting AEE by 59.4% vs RAFT256-PIV and improving generalization.]