Interpretable Kolmogorov-Arnold Networks and Symbolic Regression for Soil Compaction Parameter Prediction
Аннотация
Accurate prediction of soil compaction parameters (Optimum Moisture Content and Maximum Dry Unit Weight) is essential for earthwork infrastructure. While Artificial Neural Network (ANN) models offer high predictive accuracy, their heavily overparameterized, "black-box" architecture hinders practical deployment in geotechnical engineering, where transparent and explicit formulations are strongly preferred. This study proposes a Kolmogorov-Arnold network (KAN) framework to resolve this structural inefficiency. By integrating dynamic B-splines and L 1 sparsity, the KAN maps nonlinear, multi-phase soil behavior with greater mathematical parsimony. To ensure field applicability, Symbolic Regression (PySR) was subsequently employed to distill the KAN's optimized topology into explicit algebraic equations. Results show that the KAN model achieves comparable hold-out accuracy to the baseline ANN model while requiring 7.4-fold fewer trainable parameters (1,248 versus 9282) and maintaining phase-compliant volumetric boundaries ( S r ≤ 1.0). Benchmarking against Random Forest and XGBoost under an identical protocol showed that both mainstream ensembles matched or exceeded in-distribution accuracy, but generalized less reliably when applied to a geographically distinct Nepal database (n = 40); the KAN and PySR models instead retained their accuracy on this independent set, with the PySR equations achieving R 2 = 0.78 for both targets, a 20 = 100%, and zero physical violations across all application contexts. Out-of-distribution stress testing on highly anomalous soils (|Z|≥ 3.0) defined the explicit applicability domain. Within standard structural boundaries (LL ≤ 100%, GC ≤ 30%) under Standard Proctor compaction effort, the extracted PySR equations achieved accurate prediction ( R 2 = 0.82 for moisture; R 2 = 0.86 for density) and maintained a near-100% a 20 index. This framework bridges the gap between advanced machine learning and practical geotechnical design, providing field engineers with transparent, robust, and mathematically parsimonious compaction equations directly deployable in standard spreadsheet software.