Aplicações de redes neurais artificiais na captura de CO2: mitigação da mudança climática por processos de adsorção

Since the Industrial Revolution, atmospheric CO2 levels have increased from 280 ppm to 420 ppm by 2022, raising concerns within the scientific community about climate change. Carbon capture through gas adsorption on solid surfaces has emerged as a viable technique to address this issue. This study employed artificial neural networks (ANN) to predict CO2 uptake on activated carbon under various experimental conditions, using data such as pressure, temperature, adsorbent surface area, and uptake of CO2 and CH4. The network was trained, validated, and tested using the Levenberg-Marquardt algorithm in Matlab©, exploring architectures with 10, 15, and 20 neurons. The best performance was achieved with the 20-neuron architecture, yielding an MSE of 3.80x10⁻³ and R² values of 0.98347, 0.98328, and 0.97365 for training, validation, and testing, respectively. Additionally, the Garson method was utilized to assess the importance of the input variables, revealing that the most influential variables were surface area at 50.06%, CO2 molar fraction at 13.92%, and methane molar fraction at 13.89%. These results demonstrate the effectiveness of the ANN model in predicting CO2 adsorption, highlighting the potential of combining experimental methods with machine learning for the efficient study of greenhouse gas capture while reducing the costs and time associated with laboratory experiments.