Machine learning-enabled parameter space exploration for drilling fluid filtration

Abstract Drilling-fluid filtration is governed by coupled effects of particle size distribution (PSD), solids concentration, and particle-retention conditions, making systematic exploration by direct particle-scale simulation computationally demanding. This study integrates computational fluid dynamics–discrete element method simulations with a machine-learning surrogate to enable rapid interrogation of filtration behaviour while restricting interpretation to data-supported regions. The surrogate predicts stabilised filtrate flux from numerical retention cutoff, volume-weighted particle-size-distribution descriptors, and particle volume concentration. Predictive performance was examined using both random and input-space grouped validation, together with a k-nearest-neighbour support criterion for identifying poorly supported regions. The more demanding grouped evaluation yielded a pooled $$\:{R}^{2}=0.955$$ , demonstrating useful predictive performance when compact regions of the input space were withheld. Parameter-response analysis showed that particle volume concentration produced the strongest and most consistent response, while numerical retention cutoff and volume-weighted PSD mean both produced substantial responses; the response associated with PSD standard deviation was comparatively weaker and more condition-dependent. A targeted same-framework assessment produced a mean relative prediction error of 5.76% at selected supported conditions. The novelty of the framework lies in combining physics-based filtration simulation, support-aware surrogate interrogation, condition-dependent parameter-response analysis, and targeted numerical assessment within a single workflow. The resulting approach provides a computationally efficient screening tool for prioritising supported candidate conditions for subsequent numerical or experimental confirmation.