Data strategies for automatic calibration of soil constitutive models

Reliable numerical modeling of soil behavior depends critically on accurate calibration of representative constitutive models. While automatic calibration (AC) tools enhance efficiency and objectivity, their reliability still depends on the quality and composition of the experimental dataset. This study examines dataset composition and weighting strategies for automatic calibration of the SANISAND-F model, a sand constitutive model with evolving fabric anisotropy, using an extensive set of element tests on a low-plasticity silty sand tailings. A total of 31 tests were used, all carried out in a triaxial apparatus, with specimens consolidated under either isotropic or anisotropic conditions before shearing, including drained (CD) and undrained (CU) triaxial compression tests, drained constant- p tests (CDp), and drained K 0 triaxial compression tests (CDK 0 ). Three key questions are addressed: (1) How should experiments be weighted: by test type, or by individual test? (2) Which test combinations yield the most reliable parameters? (3) What is the minimum number and configuration of tests required for stable calibration? Targeted and constrained random datasets are assembled for parameter optimization. The optimized parameters are evaluated through full-dataset simulations and cone penetration test analyses. Results show that comparable performance is achieved under both weighting strategies, while incorporating at least two distinct test types provides sufficient complementary information to significantly enhance calibration stability. Random-subset analyses further indicate that as few as 8 appropriately selected tests may yield stable optimization results. The findings offer practical guidance for experimental design to support robust and computationally efficient constitutive-model calibration. • Automatic calibration applied to an extensive dataset of element tests on tailings. • Two error-weighting schemes evaluated for robust parameter optimization. • Calibration reliability found to depend on dataset composition and weighting scheme. • CPT simulations used to assess implications of dataset choice for model response.