Learning from theory: Rapid predictions of vibration-driven droplet motion in asymmetric pores

Seismic and acoustic vibrations can mobilize trapped non-wetting droplets, but predicting the coupled capillary–inertial response across realistic pore shapes is prohibitive at scale. We leverage a validated single-channel momentum theory to programmatically generate a high-fidelity dataset and train a physics-informed time-series neural network for time-resolved motion in asymmetric pores. Using only 3000–5000 training samples, the surrogate attains 2%–5% test error. On the same hardware, it reduces wall-clock time per case from 75 to 87 s for the mechanistic solver to less than 0.6 s, with the fastest cases completing in less than 0.3 s; this corresponds to 124–156 times acceleration and makes ensembles of thousands of pores tractable. Large surrogate ensembles reveal how frequency, peak acceleration, and geometric asymmetry govern mobilization and hysteresis, providing immediately usable constraints for models and for designing vibration strategies in enhanced recovery, groundwater remediation, and geologic carbon storage.