Simple Machine Learning Model Predicts Fluorinated Gas Capture in Ionic Liquids

Fluorinated gases are among the most potent greenhouse gases humanity has ever produced. Hydrofluorocarbons, widely used as refrigerants, carry global warming potentials thousands of times greater than carbon dioxide; HFC-23, for example, has a 100-year global warming potential of 14,800. Just as chlorofluorocarbons were banned under the Montreal Protocol for destroying ozone, HFCs are now being phased down under the Kigali Amendment. Yet capturing, separating, and recovering both legacy and modern fluorinated gases from industrial and waste streams remains a formidable technical challenge, and a new machine learning study suggests a surprisingly lean computational approach may help solve it.

Researchers have now developed a set of interpretable ensemble learning models that predict how readily fluorinated gases dissolve in ionic liquids, a class of designer solvents prized for gas separation. Ionic liquids possess negligible vapor pressure, high thermal stability, and structurally tunable chemistry, making them strong candidates for selectively absorbing refrigerants. The problem has always been scale: with more than a million possible ionic liquid structures, experimentally measuring gas solubility for every combination is financially and temporally prohibitive. Computational screening is therefore essential, and the accuracy of that screening determines whether promising solvents are found at all.

Previous machine learning efforts relied on heavyweight inputs. Artificial neural networks trained on thousands of solubility measurements required extensive sets of empirical physicochemical descriptors, including critical properties, vapor pressures, molar volumes, and fluorine atom counts, achieving average absolute relative deviations of roughly 10.93 percent. Another approach used 13 descriptors derived from COSMO-RS quantum-chemical calculations, reaching an AARD of 10.16 percent. Structure-based methods can demand as many as 47 input parameters. Such high dimensionality raises computational cost, limits practical applicability, and makes prediction difficult for newly synthesized ionic liquids whose thermodynamic properties are unknown or expensive to compute.

The new study took a deliberately minimalist route. After compiling 4,286 experimental solubility data points from the published literature and curating them with duplicate removal and interquartile-range outlier screening, the team selected just seven inputs: temperature, pressure, the critical temperature, critical pressure, and molecular weight of the ionic liquid, and the critical pressure and acentric factor of the fluorinated gas. The compiled dataset spans temperatures from 273.13 to 353.27 kelvin, pressures from 0.0095 to 3.65 megapascal, and measured solubilities from 0.0007 to 0.98 mole per mole. Spearman correlation analysis identified and removed redundant descriptors, and preliminary modeling tests confirmed that dropping correlated variables reduced overfitting risk without sacrificing accuracy. The result is a model that needs only five physically meaningful properties beyond temperature and pressure, properties that are far easier to obtain than full quantum-chemical profiles.

Four ensemble algorithms were trained and compared: CatBoost, Gradient Boosting, Random Forest, and AdaBoost with decision trees. The dataset was split 80/20 into training and testing subsets, hyperparameters were tuned by grid search with five-fold cross-validation, and the held-out test set was reserved for final evaluation only. Solubility values were logarithmically transformed to reduce skewness, and all inputs were standardized to zero mean and unit variance. CatBoost emerged as the clear winner, achieving a test mean absolute error of 0.734 percent and a test coefficient of determination of 0.991, with an overall AARD of just 3.788 percent, roughly a threefold improvement over the earlier neural network models.

The margin was not merely numerical. Bootstrap-derived 95 percent confidence intervals showed that CatBoost’s upper test mean-square-error bound of 0.028 percent fell strictly below the lower bound of 0.031 percent for the second-best model, Gradient Boosting, confirming the difference was statistically significant rather than an artifact of data splitting. The researchers attribute CatBoost’s edge to its ordered boosting scheme and symmetric, oblivious decision trees, which structurally reduce target leakage and prediction shift in correlated thermodynamic data. Random Forest and AdaBoost-DT trailed behind, with test R-squared values of 0.913 and 0.937 respectively, the latter apparently hampered by its tendency to over-focus on noisy experimental points. A learning curve for CatBoost showed training and validation losses converging and plateauing over 500 iterations without divergence, suggesting little overfitting under the adopted procedure.

Accuracy held up even at the extremes of the dataset, where machine learning models typically falter. In the low-temperature regime at or below 283.15 kelvin, relevant to evaporator and cold-absorption conditions, the model achieved an AARD of 4.39 percent. At high temperatures of 343.15 kelvin and above, corresponding to desorption stages, it reached 4.04 percent. Under vacuum pressures at or below 0.1 megapascal it delivered 5.43 percent, and at high pressures above 3.0 megapascal, where real-gas non-idealities become pronounced, it maintained 8.84 percent. Group-wise error analysis showed most predictions for individual gases and ionic liquids fell within a narrow band of plus or minus 5 percent, with only a limited number extending to plus or minus 15 percent, indicating the model’s performance was not skewed toward particular chemicals.

Crucially, the model is interpretable. SHAP analysis quantified each input’s contribution, ranking pressure first with an importance of 0.69, followed by the fluorinated gas’s critical pressure at 0.37, temperature at 0.36, ionic liquid molecular weight at 0.29, the acentric factor at 0.19, and the ionic liquid critical temperature and pressure at 0.10 and 0.05. The dominance of pressure aligns with Henry’s law, and the negative contribution of rising temperature matches the exothermic nature of physical absorption, exactly as Le Chatelier’s principle predicts. The unexpectedly high ranking of the gas’s critical pressure suggests that intrinsic condensability matters nearly as much as thermal conditions, a non-obvious insight the authors note emerges from model attribution rather than derived physical law. Local SHAP waterfall and force plots showed how individual features shift predictions from the dataset’s base value, with very low pressures of 0.01 megapascal exerting the largest negative effect in specific samples. Trend plots confirmed the model reproduces known compositional effects, such as the higher solubility of R22B1 compared with R22 and R125 in [emim][Tf₂N], attributed to bromine’s polarizability, the higher solubility of R1234yf in ionic liquids with bulky, weakly coordinating [Tf₂N]⁻ anions, and the increase in solubility with longer alkyl chains on imidazolium cations.

Leverage analysis using the Williams plot placed 97.40 percent of observations within the model’s statistically valid applicability domain, with about 1.70 percent flagged as suspected and 0.9 percent out-of-leverage. From an industrial standpoint, an overall error near 3.8 percent sits comfortably within the 5 to 10 percent uncertainty typically tolerated in conceptual absorber sizing, potentially reducing conservative over-design. As a high-throughput filter, the model can evaluate thousands of ionic liquid and fluorinated gas combinations in seconds, narrowing hundreds of candidates to a handful before any high-pressure equilibrium measurement, which can take hours to days per point, is attempted. The optimized CatBoost model is distributed as a serialized file for integration with equations of state and process simulators such as Aspen Plus.

The authors are candid about limitations. Because the dataset contains multiple measurements of the same gas-ionic liquid pairs under different conditions, the random train-test split may allow related chemical systems to appear in both subsets, so the reported performance should be read as interpolation within the existing data distribution rather than proof of extrapolation to entirely unseen chemistry. No independent external dataset was available for validation. They recommend future work using group-based validation schemes in which entire gas-solvent systems are withheld from training. Even so, the study demonstrates that a compact, physically grounded feature set paired with modern gradient boosting can outperform descriptor-heavy neural networks on the compiled data, offering a faster and more transparent path toward screening solvents for recovering the refrigerants that climate policy cannot afford to release.

Subject of Research: Interpretable ensemble machine learning for predicting fluorinated gas solubility in ionic liquids

Article Title: Interpretable machine learning for predicting fluorinated gas solubility in ionic liquids

Article References: Sheikhshoaei, A. H., Salehi, A., Naghizadeh, A., Abuswer, M. A., Schaffie, M., Mohaddespour, A., Hemmati-Sarapardeh, A., Ostadhassan, M., & Atashrouz, S. (2026). Interpretable machine learning for predicting fluorinated gas solubility in ionic liquids. Results in Engineering, 32, Article 113564. https://doi.org/10.1016/j.rineng.2026.113564

Image Credits: AI Generated

DOI: 10.1016/j.rineng.2026.113564

Keywords: fluorinated gases, ionic liquids, machine learning, CatBoost, SHAP, gas solubility, refrigerant recovery, greenhouse gases, phase equilibrium, ensemble learning, feature importance, chemical engineering