Assessing frost heave susceptibility of gravelly soils based on multivariate adaptive regression splines model

Frost heave of railway roadbed leads to track geometry degradation during the cold season, seriously threatening the safety of high-speed trains. An accurate estimation of freezing-induced deformation in subgrade aggregates is thus critical to construction and maintenance of transportation infrastructure in seasonally frozen regions. This paper proposes a practical approach to assessing the frost heave susceptibility of gravelly soils under unidirectional freezing conditions. Typical frost heave tests are first performed on gravel columns in closed and open systems. The multivariate adaptive regression splines algorithm is subsequently applied to develop a predictive model for the normalized heave of a specimen. Experimental data of freezing tests were collected from this study and available literature to compile a dataset. A randomly selected subset is used for training, while the complement of the subset is intended for testing. Relative importance analysis and analysis of variance are finally performed to examine the general and coupling effects of initial moisture content, fines, relative compaction, and stress level on the frost heave susceptibility of compacted soil. Hopefully, the developed model and comprehensive analysis of coarse fills could assist railway agencies in understanding the appropriate characterization of frost heave and provide an evaluation guideline for optimized railway roadbed.