TY  - JOUR
AU  - Calderon, Cristian J.
AU  - Tinoco, Esteban B. Hernández
PY  - 2026
DA  - 2026/08/12
TI  - On the Structural Limitations of Neural Networks for Modeling Nonlinear CO$_2$ Adsorption Equilibria
JO  - Journal of Chemical Engineering and Renewable Fuels
T2  - Journal of Chemical Engineering and Renewable Fuels
JF  - Journal of Chemical Engineering and Renewable Fuels
VL  - 2
IS  - 3
SP  - 78
EP  - 81
DO  - 10.62762/JCERF.2026.725231
UR  - https://www.icck.org/article/abs/JCERF.2026.725231
KW  - chemical engineering
KW  - neural networks
KW  - adsorption
KW  - carbon capture
AB  - Machine learning has emerged as a flexible alternative for modeling adsorption equilibrium. However, its limitations in highly nonlinear systems remain unclear. In this work, a feedforward neural network was evaluated using experimental adsorption data for CO$_2$ and N$_2$ on zeolite 13X taken from the literature. While the model accurately reproduces N$_2$ adsorption behavior, it exhibits systematic deviations for CO$_2$, particularly at higher loadings. Training diagnostics reveal a persistent validation loss plateau, indicating a limitation in the model's ability to further reduce prediction error. These results suggest that the discrepancy is not due to data quality or overfitting, but rather to a structural mismatch between the neural network representation and the strongly nonlinear, saturating nature of CO$_2$ adsorption. The findings highlight the need for hybrid approaches that incorporate physical structure, such as models based on the Langmuir isotherm.
SN  - 3070-1058
PB  - Institute of Central Computation and Knowledge
LA  - English
ER  - 

@article{Calderon2026On,
  author = {Cristian J. Calderon and Esteban B. Hernández Tinoco},
  title = {On the Structural Limitations of Neural Networks for Modeling Nonlinear CO\$\_2\$ Adsorption Equilibria},
  journal = {Journal of Chemical Engineering and Renewable Fuels},
  year = {2026},
  volume = {2},
  number = {3},
  pages = {78-81},
  doi = {10.62762/JCERF.2026.725231},
  url = {https://www.icck.org/article/abs/JCERF.2026.725231},
  abstract = {Machine learning has emerged as a flexible alternative for modeling adsorption equilibrium. However, its limitations in highly nonlinear systems remain unclear. In this work, a feedforward neural network was evaluated using experimental adsorption data for CO\$\_2\$ and N\$\_2\$ on zeolite 13X taken from the literature. While the model accurately reproduces N\$\_2\$ adsorption behavior, it exhibits systematic deviations for CO\$\_2\$, particularly at higher loadings. Training diagnostics reveal a persistent validation loss plateau, indicating a limitation in the model's ability to further reduce prediction error. These results suggest that the discrepancy is not due to data quality or overfitting, but rather to a structural mismatch between the neural network representation and the strongly nonlinear, saturating nature of CO\$\_2\$ adsorption. The findings highlight the need for hybrid approaches that incorporate physical structure, such as models based on the Langmuir isotherm.},
  keywords = {chemical engineering, neural networks, adsorption, carbon capture},
  issn = {3070-1058},
  publisher = {Institute of Central Computation and Knowledge}
}