Benchmarking Training Emissions of Regression Models for Vehicle CO2 Prediction

The urgency of climate action has intensified the use of machine learning (ML) to predict vehicular CO2 emissions; however, the training of machine learning models also generates computational emissions that are seldom reported. This study addresses a paradox central to Green AI: can carbon-intensive algorithms be justified for predicting carbon emissions? Using a public dataset of 7385 light-duty vehicles, we trained nine widely used regression models spanning simple linear baselines, polynomial and regularised linear methods, tree-based learners, ensembles, and a neural network. All experiments were instrumented with CodeCarbon to quantify real-time training footprints under a grid carbon intensity of 450 g CO2/kWh. Across models, test performance ranged from R2 = 0.72 to 0.99, yet training emissions varied by four orders of magnitude, from 0.001 g CO2 (simple linear regression) to 2.3 g CO2 (XGBoost). Although XGBoost achieved the highest accuracy (R2 = 0.9947), it emitted approximately 2300× more CO2 than regularised polynomial linear models for only a 0.39-point gain in R2. Pareto analysis identifies Lasso and Ridge regression with degree-4 polynomial features as sustainability-optimal, reaching R2 = 0.9908 at ~0.004 g CO2. To unify predictive and environmental efficiency, we introduce Accuracy-per-Gram (APG = R2/CO2) and Marginal Emissions Cost (MEC = ΔCO2/ΔR2), demonstrating a steep efficiency cliff beyond regularised linear models. At the fleet scale (100 million vehicles with daily retraining), algorithm choice implies ~84 t CO2/year for XGBoost versus ~0.15 t for Lasso, highlighting the potential climate cost of marginal accuracy gains. We provide a reproducible carbon-tracking pipeline, Green-AI evaluation metrics, and deployment guidance, arguing that computational sustainability must co-determine model selection for emissions-related ML systems. Most critically, we identify a clear accuracy–carbon emission Pareto frontier, demonstrating that regularised polynomial linear models lie on the sustainability-optimal boundary, while widely used ensemble methods such as XGBoost sit beyond an “efficiency cliff,” where marginal accuracy improvements incur disproportionately high carbon costs.