Purpose <p>Freeze-thaw cycles (FTCs), referring to the repeated alternation of soil freezing and thawing due to temperature fluctuations across 0 ℃, are a dominant geophysical phenomenon in mid- and high-latitude regions. However, their impact on carbon (C), nitrogen (N), and phosphorus (P) release in agricultural soils remains incompletely understood. This study quantifies FTC effects on soil nutrient release and develops a predictive framework for assessing nutrient loss risks under climate change.</p> Methods <p>We developed an integrated framework combining meta-analysis, machine learning, and field validation. A global meta-analysis was conducted using 1,541 datasets from 44 publications. Five machine learning algorithms were developed and validated through field experiments in Northeast China’s black soil region during the 2024–2025 freeze-thaw period.</p> Results <p>Meta-analysis revealed that FTCs significantly increased dissolved organic C by 10.37% (<i>P</i> &lt; 0.001), ammonium N by 27.35% (<i>P</i> &lt; 0.001), and available P by 11.72% (<i>P</i> &lt; 0.05). CO₂ and N₂O emissions increased by 71.94% and 307.68%, respectively (<i>P</i> &lt; 0.001). High-amplitude FTCs (≥ 30℃) enhanced C and N release by 32.71% and 54.48%. The XGBoost algorithm demonstrated optimal predictive performance (R² = 0.60, RMSE = 0.06), with field validation showing prediction deviations of only 0.18–0.65%.</p> Conclusion <p>This study quantifies FTC-driven nutrient release in agricultural soils and establishes a reliable machine learning-based prediction framework, providing a scientific basis for mitigating nutrient loss risks under future climate change.</p>