A study on optimizing a soil frost heave ratio prediction for extreme learning machines using starfish optimization algorithm

Purpose To address the challenges of extended frost heave testing cycles, potential human error in experimental measurements and the constraints of existing computationally intensive and idealized assumption-reliant frost heave theoretical models, and introduce a predictive model for soil frost heave ratio. Design/methodology/approach This study normalizes experimental frost heave ratio data from literature, demonstrated the superiority of the starfish optimization algorithm over other meta-heuristic algorithms, optimized the hyperparameters of the extreme learning machine (ELM) with this algorithm for frost heave ratio prediction, and assessed the predictive performance of the optimized SFOA-ELM model against the standard ELM model using RMSE, R2 and MAPE. Findings The SFOA-ELM model achieved an RMSE of 1.8005, an R2 of 0.9143 and an MAPE of 7.0402%, outperforming the conventional ELM model across all evaluation metrics, which substantiates the high accuracy and practical applicability of the proposed method for predicting frost heave ratios. Originality/value It innovatively introduces the Starfish Optimization Algorithm into optimizing the Soil Frost Heave Ratio prediction model for ELMs, breaking the limitations of traditional optimization methods (low efficiency and limited accuracy) and establishing a new intelligent prediction optimization path. Theoretically, it enriches the intelligent prediction method system for cold-region geotechnical engineering and improves model accuracy/stability. Practically, it provides precise data support for cold-region foundation/subgrade design, helping reduce frost heave disaster risks and offering engineering practice guidance.