Machine Learning–Based Prediction of Atmospheric CO₂ Concentration: A Year– Month Trend analysis

Israt Jahan Powsi, Rayhan Miah, Md Khorshed Alam

Atmospheric carbon dioxide (CO₂) remains the principal
driver of contemporary climate change, yet accurately
forecasting its temporal evolution requires models
capable of capturing complex nonlinear and seasonal
dynamics. In this study, we conduct a comprehensive
evaluation of thirteen supervised machine-learning
algorithms to model and predict long-term atmospheric
CO₂ concentrations using multi-decadal monthly and
annual observational datasets. The dataset encompasses
several decades of global CO₂ measurements, enabling a
detailed investigation of both persistent climatological
trends and short-term oscillatory variations.
All models were trained under a unified workflow and
assessed using a standardized performance matrix
comprising R², RMSE, and MAE. Among the tested
algorithms, Random Forest Regression and a multilayer
Artificial Neural Network (ANN) consistently
outperformed other classical and ensemble methods,
achieving R² values greater than 0.95 and demonstrating
exceptional robustness against noise and seasonal
irregularities. Time-series diagnostics further reveal a
sustained, near-exponential increase in global CO₂
levels, reflecting intensified anthropogenic influence,
reduced carbon-sink efficiency, and accelerating
feedback mechanisms in the Earth system.
The results highlight the utility of machine-learning
techniques as reliable and scalable tools for atmospheric
CO₂ forecasting, offering improved sensitivity to
nonlinearities compared to traditional statistical
approaches. Importantly, the analytical framework
developed in this work is extensible and can readily
integrate additional environmental variables such as
oceanic carbon-flux parameters, biospheric exchange
indices, remote-sensing products, or even emerging
acoustic-signature datasets to construct more holistic,
multimodal prediction systems. By establishing a strong
methodological baseline, this study contributes to the
advancement of next-generation climate-monitoring and
predictive-analytics systems, supporting ev