• Database of 1,077 test points forming 128 curves across 16 compaction energies • PCA-KNN selects similar historical tests as neighbors • GPR reconstructs compaction curves and posterior bands from a single observation • Compaction is governed mainly by energy and moisture-related variables • 89.1% of measured peak dry densities fall within the model-derived posterior intervals Subgrade compaction is a critical determinant of road pavement performance and durability. Conventional quality control relies on laboratory compaction curves to determine the maximum dry density (MDD) and optimum moisture content (OMC), but generating a full curve requires multiple test points, and the peak can be interpreted inconsistently when the curve becomes weakly concave-down, near-linear, or concave-up at low compaction energy. Using a laboratory database containing 1,077 test points from 128 curves for 0-2.36 mm crushed sandstone aggregates, we first assessed how well the dataset supports scalar prediction of MDD/OMC. Multiple linear regression yielded R² = 0.319 for MDD, while gradient-boosted regression trees achieved R² ≈ 0.77 for both MDD and OMC. However, these scalar models provide point estimates only and do not reconstruct the full moisture-dry density relationship needed for curve-based interpretation and targeted additional testing under sparse field observations. We therefore present a data-efficient workflow that reconstructs the compaction curve and extracts MDD/OMC from as few as one moisture-density observation by combining (i) PCA-based similarity mapping, (ii) KNN retrieval of comparable historical tests, and (iii) Gaussian process regression (GPR) to obtain a smooth mean curve with model-derived posterior uncertainty. Across the full dataset, MDD predictions extracted from the reconstructed curves yields MAE = 0.056 Mg/m³, MAPE = 3.1% ( R² = 0.681), and 89.1% of measured MDD values fell within the model-derived posterior bands. The applicability of the reported results is limited to the material family and test method represented in this paper.
Kamura et al. (Wed,) studied this question.