Granum, an agentic-AI GPU discrete element program: verification for dynamic granular flows

Granum, a GPU discrete element program for soil that an AI agent can drive through theModel Context Protocol, written for quasi-static tests and staged models, was extended fordynamic granular flow with hopper and silo walls, a turning drum, free-falling rigid bodies,grain sinks and counters, initial states read from files and pair-specific contact properties, andwas then tested at three levels. At the contact level it reproduces the impact benchmarksof Chung and Ooi to within 0.1% of Hertz’s solutions. In the DEM8 benchmark of nineopen-source codes it falls inside the codes’ band in all three bulk processes: the four silo flowrates lie within 0.7 to 1.9% of the middle of the band and the ratio of large to small orifice is3.10 against 3.10 ± 0.06, the number of grains in a quadrant of the rotating drum follows thecodes within 246 grains rms with its peaks within 0.04 s and 2%, and the impact paths liewithin the codes’ spread. Against laboratory laws, the discharge of a sand through orifices of5 to 10 grain diameters follows Beverloo’s law with C = 0.54 and k = 1.51, and the collapseof granular columns with aspect ratios of 0.5 to 8 matches the runout of Lube et al. and thefinal height of Lajeunesse et al. once the sand’s shape and rolling resistance are calibratedon its angle of repose. A steel ball penetrates 0.72 to 0.94 times the depth of Uehara etal.’s law, with a stopping time nearly independent of the impact speed as Katsuragi andDurian found, and a 10 t tamper dropped from 5 to 15 m onto a 16 m loose fill reaches adepth of influence of 0.47 to 0.59√WH, within the field range of Mayne et al. The testsalso exposed three defects, a braking effect of fixed drum end plates, a placement step thatignored an impactor’s width, and an all-pairs contact search caused by one large sphere,which are reported with their fixes.