Physics-informed neural network for modeling force and torque fluctuations in a random array of bidisperse spheres

Abstract: We present a physics-informed neural network (PINN) model to predict the hydrodynamic force and torque fluctuations in a random array of stationary bidisperse spheres. The PINN model is formulated based on two hypotheses: (i) pairwise interaction assumption that approximates the total force/torque exerted on a target sphere by linear superposition of individual contributions from a finite number of influential neighbors; (ii) unified function representation that suggests a single functional form to describe the contribution from different neighbors based on the observation of probability distribution maps obtained with various binary interaction modes in bidisperse particle-laden flows. On this basis, we accordingly establish a compact PINN architecture to evaluate individual force/torque contribution of influential neighbors through the same neural network block which tremendously reduces the number of unknown parameters, and ultimately compute the total force/torque exerted on target sphere by their weighted sum. We compare the model predictions to Particle-Resolved Direct Numerical Simulation (PR-DNS) data of eight different cases in a range of Reynolds number $1\leq Re\leq100$, total solid volume fraction $10\%\leq\phi\leq40\%$, sphere diameter ratio $1.5\leq d_{l}^{*} /d_{s}^{*}\leq2.5$ and volume ratio $1\leq V_{l}^{*}/V_{s}^{*}\leq 4$, which demonstrates excellent performance with an optimal $R^2\approx0.9$ for both force and torque predictions. We establish a universal model that is applicable within the aforementioned input space, and examine its interpolation capability to the unseen data with multiple additional datasets. Finally, we extract and illustrate the interpretable information of our PINN model in binary and trinary interactions, and discuss its potential extensions to other particle-laden flow problems with more complicated scenarios and Eulerian-Lagrangian simulations as a superior alternative to the classic average drag laws.