Abstract

Calibrating the Lattice Discrete Particle Model (LDPM), a mesoscale framework for simulating concrete, is computationally and experimentally demanding because it requires data across multiple material scales. This study examines how machine learning (ML)-based concrete compressive strength prediction can support LDPM-oriented multiscale concrete modeling. While ML has been widely applied to predict concrete strength from mixture proportions, its role as a benchmarking and diagnostic tool for physically based multiscale modeling remains less established. Six regression models were trained using the 1030-sample concrete compressive strength dataset originally compiled by Yeh, with concrete mixture components and curing age used as input variables. The models were evaluated through train–test validation, five-fold cross-validation, feature importance analysis, application to an experimentally validated LDPM mixture, and external validation using additional LDPM reference mixtures. The results show that tree-based ensemble models provided the strongest predictive performance, with Extra Trees (ET) achieving an RMSE of 4.8 MPa. When applied to the LDPM validation mixture, the ML predictions showed close agreement with the low-friction (LF) compressive strength reference, consistent with the low-confinement conditions represented in standard compressive strength Yeh’s datasets. This agreement supports using the LF value as the primary benchmark for the present ML–LDPM comparison. Because the ML inputs do not include specimen geometry, boundary friction, full aggregate gradation, or LDPM-specific mesoscale parameters, the predictions should be interpreted as preliminary strength benchmarks rather than substitutes for experimental testing or detailed LDPM calibration. External validation further showed that ML models can provide useful estimates for LDPM-relevant mixtures, particularly when the target mixtures fall within or near the training-data range. Overall, the study demonstrates that large experimental databases and accessible ML tools can support LDPM-oriented workflows by providing rapid preliminary screening, benchmarking, and diagnostic interpretation while preserving the need for physically based LDPM validation and calibration.

IPC Classification

Keywords