HyMARC Sorbent ML

Predicting the hydrogen storage capacity of metal-organic frameworks via machine learning

This webpage presents computational models for predicting the usable hydrogen capacity of metal-organic frameworks (MOFs). The models are based on machine learning (ML) algorithms that have been trained on computed H2 uptake data (i.e., isotherms from Grand Canonical Monte Carlo calculations) on a diverse set of MOFs. A full description of the models and training procedure is coming soon.

Capacities are predicted for two operating conditions:

An isothermal pressure swing (PS) between 100 and 5 bar at T = 77 K
2. A temperature + pressure swing (TPS) between P = 100 bar / T = 77 K (full condition) and P = 5 bar / 160 K (empty condition)

Additionally, for each operating condition two types of capacities are reported:

Usable gravimetric capacity in weight percent (wt.%), defined as the mass of stored H2 divided by the summed mass of H2 and the MOF
Usable volumetric capacity in grams of H2 stored per liter of MOF (g H2/LMOF)

The predictions of volumetric capacity assume that the MOF is a single crystal (monolith). Hence, losses due to imperfect packing (i.e., void space) that might be present in a storage system where the MOF is stored as a powder are not accounted for. The present volumetric predictions should therefore be considered as an upper bound on the practical volumetric capacity of the MOF storage medium.

The capacity predictions require crystallographic data from the user. These input data are referred to as ‘features.’ The seven features (and their allowable ranges) that can be used by the ML models are:

Single crystal density, d (g cm-3); range = 0.03 – 5.18
Gravimetric surface area, gsa (m2 g-1); range = 0 – 9750
Volumetric surface area, vsa (m2 cm-3); range = 0 – 3995
Void fraction, vf (unitless, range of 0-1); range = 0 – 0.99
Pore volume, pv (cm3 g-1); range = 0 – 35.7
Largest cavity diameter, lcd (Å); range = 0.4 – 71.6
Pore limiting diameter, pld (Å); range = 0 – 71.5

Predictions can be made using as few as one feature or with a maximum of all 7. (Specifically, models requiring input of 1, 4, 5, and 7 features are available.) In general, the more features that are used, the more accurate are the predictions. Table 1 summarizes the accuracy of the models as a function of the number of features, capacity type, and operating condition. Input values outside of the feature ranges listed above should not be used.

Table 1. Performance of ML models for the prediction of usable H2 storage capacities in MOFs under pressure swing (PS) and temperature + pressure swing (TPS) conditions. Here d, gsa, vsa, vf, pv, lcd, pld represent single crystal density (g cm-3), gravimetric surface area (m2 g-1) , volumetric surface area (m2 cm-3), void fraction, pore volume (cm3 g-1), largest cavity diameter (Å), and pore limiting diameter (Å), respectively. Performance is assessed based on R2, average unsigned error (AUE), and root-mean-square-error (RMSE).

Values for the features can be measured experimentally, computed from the MOF’s crystal structure, or, for the adventurous, simply guessed at. If one has access to the crystal structure then the zeo++ code can be used to quickly compute these features from a MOF’s crystallographic information file (CIF).

Acknowledgement

Financial support for this work was provided by the US Department of Energy, Office of Energy Efficiency and Renewable Energy, Grant no. DE-EE0007046. Partial computing resources were provided by the NSF via grant 1531752 MRI: Acquisition of Conflux, A Novel Platform for Data-Driven Computational Physics (Tech. Monitor: Ed Walker). If you find these tools helpful in your research, or if you publish predictions made using these models, we would be grateful if the following article could be cited: A. Ahmed and D. J. Siegel, Predicting Hydrogen Storage in MOFs via Machine Learning, Patterns, 2, 100291 (2021). DOI: 10.1016/j.patter.2021.100291