Deep learning‐assisted <scp>DEM</scp> simulation for optimizing concrete mix proportions

Abstract This study utilized the DECAS‐Net model to precisely segment concrete images containing complex aggregate distributions (mIoU = 85.90%), enabling the construction of a comprehensive database of realistic aggregate geometries. Based on this database, two types of numerical models were developed: concrete models incorporating real aggregates with varying volume fractions and corresponding PFC2D models using circular aggregates for comparative analysis. The systematic investigation was conducted to examine the influence of loading rate, aggregate volume fraction, and aggregate shape on both the macroscopic and microscopic mechanical responses of the concrete. The results demonstrated that, under quasi‐static loading conditions, the model incorporating real aggregate geometries combined with the Bonded Particle Model (BPM) yielded reliable performance, with an optimal loading rate threshold identified at 0.08 m/s. Furthermore, it was observed that appropriately increasing the aggregate volume fraction generally enhances mechanical properties, with the optimal range identified between 5% and 10%, where the concrete exhibits peak compressive strength and Young's modulus. These findings directly inform the optimization of concrete mix proportions in engineering design, enabling the development of high‐strength, durable structures with tailored aggregate distributions.