Machine-learning models for predicting CO2 solubility in various brine systems: implications for carbon geo-storage
| Parent link: | Journal of Molecular Liquids.— .— Amsterdam: Elsevier Science Publishing Company Inc. Vol. 435.— 2025.— Article number 128122, 24 p. |
|---|---|
| その他の著者: | , , , , , |
| 要約: | Title screen Accurate prediction of CO2 solubility in brine is critical for ensuring the safe and efficient storage of CO2 and assessing the long-term stability of storage sites. Existing models often lack generalizability and are limited to specific brine systems. This study addresses these limitations by developing hybrid models that utilize Long Short-Term Memory (LSTM) networks in conjunction with five metaheuristic optimization algorithms to predict CO2 solubility (mol/kg) across diverse brine systems. The main input variables include temperature (K), pressure (MPa), and concentration (mol/kg) of nine salts (NaCl, KCl, CaCl2, MgCl2, MgSO4, Na2SO4, NaHCO3, K2SO4, and K2CO3). A comprehensive laboratory dataset was preprocessed by dividing the data into training (70 %), validation (15 %), and test (15 %) sets, followed by normalization and outlier detection using the leverage technique, which identified and removed nine outliers. Before the models were applied to the training data, the LSTM structure and controllable parameters were optimized using metaheuristic optimizers. To ensure model reproducibility, the prediction of each algorithm was run five times. The hybrid models demonstrated lower mean and standard deviation of RMSE than the standalone LSTM, indicating higher stability in prediction performance. The best-performing model, LSTM-COA (cuckoo optimization algorithm) hybrid, achieves the lowest RMSE values: 0.0339, 0.0523, and 0.0632 mol/kg during training, validation, and testing phases, respectively. The bootstrapping analysis confirmed that hybrid models, particularly LSTM-COA, were less susceptible to overfitting. This result demonstrates better generalization to real-world data. Furthermore, uncertainty assessment was used to validate the reliability of hybrid models, with LSTM-COA showing an excellent performance in predicting CO2 solubility. The LSTM-COA model demonstrated robust performance for pure water, most single-salt, and mixed-salt systems, particularly MgSO4 and NaCl-KCl, but struggled with NaHCO3-containing solutions and complex mixed systems involving NaCl, KCl, CaCl2, MgCl2, and NaHCO3. Feature importance analysis via SHAP identified temperature as the most influential input, while the concentration of NaHCO3 had the least impact. This study highlights the effectiveness of hybrid LSTM models, with LSTM-COA emerging as a precise and reliable tool for CO2 solubility prediction in carbon geo-storage projects, offering practical applications for optimizing CO2 storage site selection and monitoring Текстовый файл AM_Agreement |
| 言語: | 英語 |
| 出版事項: |
2025
|
| 主題: | |
| オンライン･アクセス: | https://doi.org/10.1016/j.molliq.2025.128122 |
| フォーマット: | 電子媒体 図書の章 |
| KOHA link: | https://koha.lib.tpu.ru/cgi-bin/koha/opac-detail.pl?biblionumber=681632 |