Advanced AI model pushes aviation toward data-driven decarbonization

Researchers have developed an advanced artificial intelligence (AI) framework designed to significantly improve the forecasting of carbon dioxide emissions in the aviation sector.

Their study, titled AI-Driven Predictive Analytics for Sustainable Aviation: Metaheuristic-Optimized XGBoost for Carbon Emission Prediction and published in the journal Sustainability, introduces a hybrid machine learning architecture that combines Extreme Gradient Boosting with an enhanced metaheuristic optimizer known as ACGRIME, demonstrating measurable gains in predictive accuracy for aviation-related CO₂ emissions.

Reinventing hyperparameter optimization with ACGRIME

ACGRIME is an improved metaheuristic algorithm derived from the original RIME framework. ACGRIME integrates three strategic mechanisms: chaotic initialization, adaptive weighting and Gaussian mutation. Together, these components enhance search diversity, stabilize convergence behavior and prevent entrapment in local optima.

Chaotic initialization expands the global search capacity during early iterations, reducing the risk of premature convergence. Adaptive weighting dynamically balances exploration and exploitation as optimization progresses, ensuring that the algorithm initially scans broad solution spaces before refining promising regions. Gaussian mutation introduces stochastic variations that preserve population diversity and improve robustness.

By integrating ACGRIME into XGBoost’s hyperparameter tuning process, the researchers created a hybrid framework capable of identifying optimal learning rates, tree depths and boosting configurations more effectively than conventional tuning methods. The hyperparameter space explored included learning rates between 0.01 and 0.3, boosting tree counts from 50 to 200 and tree depths between 3 and 10. These parameters are central to controlling model complexity, convergence speed and predictive performance.

To validate ACGRIME’s optimization capabilities, the authors benchmarked it against several established nature-inspired algorithms using the CEC 2020 optimization suite. Nonparametric statistical testing, including Wilcoxon Rank-Sum and Friedman tests, confirmed that ACGRIME consistently achieved superior convergence performance and lower average ranking scores across multiple benchmark functions. The results reinforced the optimizer’s robustness and generalizability beyond a single predictive task.

This benchmarking phase was vital. Rather than applying the hybrid model directly to aviation data, the researchers first demonstrated that ACGRIME itself was statistically competitive across diverse optimization problems. Only after establishing this performance foundation did they proceed to real-world emissions forecasting.

Applying hybrid AI to aviation CO₂ forecasting

The evaluation focused on a dataset of international flight records sourced from Kaggle. The dataset contained operational variables including origin country, destination country, aircraft type, airline number, flight duration, number of stopovers, ticket price and estimated CO₂ emissions per flight. Preprocessing steps ensured data integrity, including encoding categorical variables, standardizing numerical features and removing missing values.

The dataset was divided into training and testing subsets using an 80:20 ratio to ensure rigorous assessment of generalization performance. Model accuracy was evaluated using multiple widely accepted performance metrics: Mean Squared Error, Root Mean Squared Error, Normalized RMSE, Mean Absolute Error and the coefficient of determination, R².

The ACGRIME-XGBoost model achieved a test R² score of approximately 0.94, outperforming baseline XGBoost and other metaheuristic-enhanced variants. Error metrics were significantly lower than those recorded by competing models, indicating stronger predictive precision and reduced deviation between observed and predicted CO₂ values. Training performance approached near-perfect fit while maintaining strong testing accuracy, signaling controlled generalization rather than overfitting.

Several comparative models were assessed, including RIME-XGBoost, EDO-XGBoost, HHO-XGBoost, PLO-XGBoost, SCA-XGBoost, SCSO-XGBoost and TSO-XGBoost. While some of these demonstrated competitive performance, none matched the consistent convergence stability and accuracy achieved by ACGRIME-XGBoost.

Feature importance analysis yielded another critical insight. Destination country emerged as the most influential predictor of CO₂ emissions, far outweighing other operational variables such as number of stops, ticket price, aircraft type or duration. This suggests that route distance, regional airspace regulations and geographic characteristics may exert dominant influence over emission outcomes. The finding reinforces the complexity of aviation carbon modeling and highlights the importance of incorporating spatial context into predictive systems.

Implications for sustainable aviation and smart infrastructure

ACGRIME-XGBoost is more than a technical advancement. The hybrid model is a decision-support tool capable of contributing to sustainable smart infrastructure development. Accurate carbon emission forecasting can inform route optimization, fleet management decisions, regulatory planning and carbon offset strategies.

As aviation faces pressure to align with global net-zero targets by 2050, predictive analytics will play an increasingly major role. Real-time integration of AI-driven emission forecasts with Internet of Things sensor networks and digital monitoring systems could enhance environmental transparency and policy enforcement.

However, the authors also acknowledge several limitations.

• The aviation dataset used for validation primarily included flights originating from China, which may limit global generalizability. Operationally significant variables such as aircraft payload, cabin configuration, cruising altitude, engine type and airline-specific fuel management practices were not included due to data constraints. These factors could further refine predictive performance if incorporated in future research.
• Computational cost represents another consideration. Metaheuristic optimization algorithms can require significant processing resources, particularly in large-scale or real-time applications. 
• The absence of ablation analysis isolating the individual contributions of chaotic initialization, adaptive weighting and Gaussian mutation was also noted.

Future research directions include parallelization strategies, development of lightweight optimization variants and integration of explainable AI techniques to enhance transparency. The authors propose conducting comprehensive ablation experiments in future studies to quantify the impact of each enhancement within ACGRIME.