AI Model Predicts Methane Storage in Metal-Organic Frameworks

Researchers have developed a machine learning surrogate capable of predicting methane deliverable capacity across a database of metal-organic frameworks with high accuracy. The study, published in Discover Artificial Intelligence, evaluates how well a descriptor-based model can reproduce simulated methane deliverable capacity and whether it can be inverted to design new materials. The team analyzed a public database containing 252,353 frameworks, using a histogram-based gradient boosting regressor to map structural and chemical descriptors to simulated uptake. The model achieved a coefficient of determination of 0.9963 and a mean absolute error of 0.329 millimoles per gram under a random split, demonstrating strong predictive power for the specific dataset used.

The database is composed of 97.58% computationally enumerated frameworks and only 2.42% derived from experimentally reported crystal structures. This composition significantly influences the model’s performance, as the enumerated frameworks occupy a distinct region of property space compared to the experimental subset. The enumerated frameworks average a deliverable capacity of 11.69 millimoles per gram at a mean density of 0.67 grams per cubic centimeter, whereas the experimentally derived subset averages 2.79 millimoles per gram at 1.31 grams per cubic centimeter. This imbalance means the model primarily learns the internal regularity of a small number of generative procedures rather than the full diversity of realizable materials.

To ensure the model’s accuracy was not an artifact of database redundancy, the researchers subjected it to eight structure-aware validation protocols. These included grouping by net topology, metal-node class, and organic-linker family to prevent near-duplicates from appearing on both sides of the train-test split. Under the most demanding structural constraint, grouping by topology, the model’s coefficient of determination fell to 0.9877, and the mean absolute error rose by 73% to 0.568 millimoles per gram. Despite these changes, the model maintained high accuracy, indicating that the predictions are driven by learned relationships rather than simple lookup of duplicate entries.

A critical limitation emerged when the model was trained exclusively on enumerated frameworks and tested on the 6,108 experimentally derived ones. In this cross-population test, the Spearman rank correlation dropped from 0.998 to 0.879, indicating a marked reduction in the model’s ability to order experimental frameworks correctly. This result suggests that while the surrogate is highly accurate for the enumerated population, its performance on experimentally realizable materials is lower. The study highlights the importance of considering the provenance of data when evaluating machine learning models for materials discovery.

The researchers also attached distribution-free split-conformal prediction intervals to every prediction to provide a measure of uncertainty. In the top 1% of frameworks by simulated capacity, the 90% intervals covered only 74.7% of the true values. This failure of conditional coverage demonstrates that marginal calibration does not guarantee reliability for individual high-performing candidates, a key consideration for prioritizing expensive downstream synthesis.

When benchmarked against direct screening, the descriptor-space optimization approach did not outperform simple ranking of the database. Direct model screening recovered the database record holder, a framework with a deliverable capacity of 57.89 millimoles per gram, in a top-ten shortlist. In contrast, the optimise-then-retrieve method returned frameworks with capacities ranging from 17.7 to 28.8 millimoles per gram and recovered none of the true top ten candidates. This negative result indicates that in a fully labelled database, ranking the existing entries is more effective than attempting to design new descriptor vectors.

The optimized descriptor vectors produced by the model were found to be physically unattainable. The optimal fingerprint violated the rigid-framework mass-balance identity, where void fraction equals the product of density and gravimetric pore volume. The residual for the optimized vector was several orders of magnitude larger than the standard deviation observed in real frameworks. Furthermore, no real framework in the database matched all eight optimized descriptors simultaneously, even with a 50% relative tolerance. This confirms that the mathematical optimum lies outside the manifold of physically possible structures.

Despite the negative results for inverse design, the study provides a validated surrogate that reduces the cost of scoring the entire database from months of simulation to seconds. The model identifies that 14 geometric descriptors are sufficient for regression, with the remaining 148 chemical descriptors primarily improving ranking quality at the head of the list. Crystallographic density and probe-occupiable gravimetric pore volume were identified as the most influential descriptors, consistent with the physical intuition that accessible pore volume and framework mass determine methane uptake. The study recommends the 14-descriptor geometric model as the default for gravimetric screening.

No framework in the database meets the U.S. Department of Energy’s volumetric target of 263 cubic centimeters of standard temperature and pressure methane per cubic centimeter of adsorbent. The maximum single-crystal volumetric capacity in the database is 213.5 cubic centimeters per cubic centimeter, which falls to 160.1 at 75% packing. However, 1,405 frameworks, or 0.56% of the database, exceed the gravimetric target of 0.5 grams per gram. The best gravimetric performer reaches 0.929 grams per gram, but its low density results in a volumetric capacity of only 30% of the target. This highlights the conflict between gravimetric and volumetric objectives in methane storage material design.

The study concludes that while machine learning surrogates can accurately predict simulated properties, they do not currently accelerate discovery in labelled databases. The retrieved candidates are database-enumerated structures with no experimental validation, and the optimized fingerprints are not physically realizable. Future work should focus on coupling the latent-space design formulation with generative models that propose new structures and validating high-ranking experimentally derived frameworks against measured isotherms. The study provides a rigorous framework for assessing the reliability and limitations of machine learning in materials science.

The study’s reliance on simulated data rather than experimental measurements is a central constraint on its applicability. All labels in the database are outputs from grand canonical Monte Carlo simulations conducted at 5.8 and 65 bar, the pressure window adopted by the U.S. Department of Energy for onboard methane storage. The researchers did not recompute these labels but used them as published, meaning the surrogate inherits the assumptions of the original force fields and rigid-framework treatments. Consequently, the model’s accuracy reflects its ability to reproduce these specific computational outputs, not its capacity to predict measured uptake in real-world conditions. The authors explicitly state that no experimental validation was performed, and the database provides no measured isotherms to benchmark against. This distinction is critical because the high predictive performance observed in the study cannot be directly translated into confidence in the model’s utility for guiding the synthesis of new materials without independent experimental verification.

Previous work in this field has often focused on predicting methane uptake at a single pressure or reporting point predictions without calibrated uncertainty estimates. This study differentiates itself by modeling the deliverable capacity directly, defined as the difference between the loading at the charging pressure and the loading at the depletion pressure. This approach aligns the machine learning objective with the engineering requirement, as only the difference in loading reaches the engine during operation. By targeting this working capacity rather than single-point uptake, the researchers address a specific gap in prior literature, which frequently operated purely in a forward-ranking mode without inverting the model for design purposes. The methodological focus on the working capacity provides a more relevant metric for evaluating the practical viability of adsorbed natural gas storage systems.

The conflict between gravimetric and volumetric storage targets is further illustrated by the lack of overlap between the top performers in each category. The study found that the Spearman correlation between the two objectives across the database was 0.780, and the overlap between the top 1% of frameworks under each objective was zero. No framework that ranked among the best 2,524 gravimetrically was also among the best 2,524 volumetrically. This disjointness indicates that optimizing for one metric does not yield progress on the other. The record gravimetric holder, with a density of 0.061 grams per cubic centimeter, achieves only 78.9 cubic centimeters per cubic centimeter, or 30% of the volumetric target, because of its low density. This structural trade-off suggests that future materials design must explicitly balance these competing objectives rather than treating them as interchangeable measures of performance.

The study also examined the physical plausibility of the optimized descriptor vectors by checking them against the manifold of real frameworks. The optimal fingerprint was found to be an outlier, with a Mahalanobis distance of 26.3 in the principal-component space of the real frameworks. This distance is significantly larger than that of the retrieved candidates, which ranged from 2.6 to 4.3. The isolation forest fitted on the real frameworks flagged the optimum as an outlier, confirming that the combination of descriptor values proposed by the model does not correspond to any physically realizable structure. This unphysical outcome occurs because the optimizer holds strongly coupled geometric descriptors fixed at their medians while moving others, violating physical identities such as the framework mass-balance relation.

Despite the negative results for inverse design, the study offers a validated surrogate that can significantly reduce the computational cost of screening large databases. Replacing per-framework simulation with this model reduces the time required to score the entire database from months to seconds. This speed-up is a practical benefit for researchers who need to evaluate large numbers of candidates quickly. The authors emphasize that influential descriptors like crystallographic density and probe-occupiable gravimetric pore volume indicate what the model relies on, rather than separate variables that physically govern methane uptake. These descriptors are closely related to the accessible free volume per unit mass, which is a key determinant of adsorption capacity. The study’s findings suggest that a smaller set of geometric descriptors may be sufficient for initial screening, reserving the full chemical descriptor set for applications where the ranking of top candidates is critical.

Subject of Research: Machine learning surrogate for methane storage in metal-organic frameworks

Article Title: Reliable methane storage screening and inverse design of metal organic frameworks under distribution shift

Article References: Chowdhury, S. H., Diptho, R. A., & Ghosh, P. (2026). Reliable methane storage screening and inverse design of metal organic frameworks under distribution shift. Discover Artificial Intelligence, 6(1), Article 1726. https://doi.org/10.1007/s44163-026-02574-5

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02574-5

Keywords: metal-organic frameworks, methane storage, machine learning, gradient boosting, conformal prediction, inverse design, adsorbed natural gas, materials discovery, Reliable, methane, storage, screening