UC Riverside

The deleterious impact of erosion due to high-velocity particle impingement adversely affects a variety of engineering and industrial systems, resulting in irreversible mechanical wear of materials/components. Brute force computational fluid dynamics (CFD) calculations are commonly used to predict surface erosion by directly solving the Navier-Stokes equations for fluid and particle dynamics; however, these numerical approaches often require significant computational resources. In contrast, recent data-driven approaches using machine learning (ML) have shown immense promise for more efficient and accurate predictions to sidestep computationally demanding CFD calculations. To this end, we have developed FLUID-GPT (Fast Learning to Understand and Investigate Dynamics with a Generative Pre-Trained Transformer), a new hybrid ML architecture for accurately predicting particle trajectories and erosion on an industrial-scale steam header geometry. Our FLUID-GPT approach utilizes a Generative Pre-Trained Transformer 2 (GPT-2) with a convolutional neural network (CNN) for the first time to predict surface erosion using only information from five initial conditions: particle size, main-inlet speed, main-inlet pressure, subinlet speed, and subinlet pressure. Compared to the bidirectional long- and short-term memory (BiLSTM) ML techniques used in previous work, our FLUID-GPT model is much more accurate (a 54% decrease in the mean squared error) and efficient (70% less training time). Our work demonstrates that FLUID-GPT is an accurate and efficient ML approach for predicting time-series trajectories and their subsequent spatial erosion patterns in these complex dynamic systems.