Physics-Informed Data Augmentation of Experimental Erosion Data Through Generative Adversarial Networks

Abstract Erosion is a phenomenon of material removal by particles impinging the walls of materials. Experiment is an important approach to investigate erosion severity and the influences of different physical parameters. However, conducting experiments is expensive and time-consuming. It’s also challenging to collect high quality data due to the scales of the material removal and its sensitivity to the change of external environments. Over the years, research has been conducted at the Erosion/Corrosion Research Center at the University of Tulsa to investigate solid particle erosion in pipelines and fittings. The experiments are conducted with a large-scale boom tower flow loop and erosion data is collected with high accuracy temperature compensated ultrasonic transducers. The efforts have resulted in more than 200 sets of high-quality experimental data under various pipe diameters, particle diameters, sand concentrations and flow velocities. Mining the inherent pattern and/or laws within the data is of great value to guide future process condition selections and experimental tests. More importantly, with the existing data that has covered a wide range of flow, particle, pipe size and sand concentration conditions, machine learning methods can be utilized to augment the data itself so that the new generated data obeys the physical laws and have similar distributions. Generative Adversarial Networks (GANs) is a class of machine learning models designed for generating data. It has a broad application. Data augmentation is an area where it can be applied to generate additional training data when data is limited for later machine learning or deep learning. The feature is of significance to expand erosion databank based on a relatively small high-quality database for subsequent mechanistic modeling. This present work shows how to apply GANs to achieve the goal and the newly generated data are validated with both original data and the data obtained through Computational Fluid Dynamics of which physics of unseen conditions are fully simulated and SPPS which is a semi-mechanistic model developed at E/CRC. The work provides engineers and researchers with a new approach to expand databases and build mechanistic models upon it.