Generalization of a surrogate drag force model for particle laden flowVovk, Nejc (Avtor) Kokovnik, Jakob (Avtor) Ravnik, Jure (Avtor) multiphase flowparticle-laden flowparticle-resolved simulationdrag forcesurrogate modelsmachine learningWe present a three-way validation of a data-driven interparticle force model against high-fidelity direct numerical simulations (DNS) and classical Faxén predictions, in the context of laminar particle-laden pipe flow. We present an Interaction-Decomposed Neural Network (IDNN) architecture, trained for unbounded flows, and perform finite-volume DNS that fully resolve the flow around stationary spherical particles placed in a circular pipe. This allows direct measurement of hydrodynamic forces for a range of particle configurations and flow conditions. Forces computed from the DNS are compared to the IDNN model evaluated on the same instantaneous geometric inputs and to analytical estimates from Faxén-type expansions. Although the IDNN was trained on unbounded flows, we show it remains predictive in pipe geometries provided the nearest walls are sufficiently distant compared with particle cloud size; in these cases the IDNN reproduces the DNS force components and trends with particle spacing and relative orientation with close quantitative agreement. The three-way comparison confirms that the IDNN can accurately reproduce resolved forces while avoiding the high cost of particle-resolved simulations and the complex Laplacian evaluations required by Faxén theory. These results inform the use of data-driven force models in Euler–Lagrange simulations and guide future model improvement for wall-influenced and strongly interacting particle systems.Elsevier20262026-06-04 14:16:07Članek v reviji98315sl