Resolving multi-objective conflicts in earth-air heat exchanger (EAHE): A prediction-analysis-optimization framework

The Earth-Air Heat Exchanger (EAHE) is a promising renewable energy technology, yet its engineering application is hindered by trade-offs between thermal performance, energy efficiency, and economic viability. To address this, this study proposes a novel framework integrating a mathematical model, machine learning (ML) and multi-objective optimization (MOO) for EAHE performance prediction, analysis, and optimization. Initially, a validated mathematical EAHE model was employed to generate 10,000 datasets for each of three Chinese cities with distinct climates, which were then used to develop ML-based performance prediction models. Subsequently, spearman's rank correlation coefficient and SHapley Additive exPlanations (SHAP) were applied to reveal complex relationships between performance indicators and design parameters, quantifying regional differences and similarities in mechanisms. Finally, nine MOO algorithms were used to generate optimization solutions for the three cities under different scenarios. The results indicate that: (1) the XGBoost regressor demonstrated superior accuracy in predicting EAHE performance across all regions (R2 > 0.97, RMSE ≈ 26), outperforming other ML models; (2) complex correlations exist between performance metrics, with none simultaneously achieving ideal ranges, although the influence patterns of design parameters on performance are similar across regions; and (3) among nine MOO algorithms, NSGA-II-DE achieved optimal Pareto solutions (500 non-dominated solutions, hypervolume >0.98) within 2 min. Compared to mathematical model-driven optimization, XGBoost-driven NSGA-II-DE optimization is 500 times more efficient. Scenario-specific decision-making provided six tailored optimization schemes per region. The findings support the design optimization of EAHE and accelerate the realization of sustainable development goals.