Attainable Volumetric Targets for Adsorption-Based Hydrogen Storage in Porous Crystals: Molecular Simulation and Machine Learning

Hydrogen
fuel is attractive to power vehicles without emitting
carbon, but onboard storage of sufficiently densified hydrogen at
moderate pressure remains a significant challenge. Adsorption-based
storage in porous crystals such as metal–organic frameworks
and covalent organic frameworks is attractive to reduce the storage
pressure. It is, however, unclear to what extent volumetric storage
targets can be met under constraints of adsorbent design and choice
of operating conditions. To help elucidate attainable values for volumetric
storage metrics upon the potential introduction of strong hydrogen-binding
sites, we “computationally synthesized” a library of
porous crystals and performed 18 000+ grand canonical Monte
Carlo simulations to calculate hydrogen loadings at multiple T, P conditions. The studied frameworks
are based on 17 pore topologies and feature alchemical catecholate sites: sites whose interaction with hydrogen was artificially
and systematically modified within the range of density functional
theory-calculated hydrogen–catecholate binding energies found
in the literature. Porous crystals with the tetrahedrally connected
topologies dia and qtz tended to outperform
other types of crystals for each “level” of hydrogen–alchemical
site interaction strength. Among tested operating conditions, 100
bar/77 K ↔ 5 bar/160 K swing conditions produced the highest
optimal deliverable capacity (62 g/L with a 10 kJ/mol heat of adsorption),
which was 138% higher than that for the 100 bar ↔ 5 bar swing
at ambient temperature (26 g/L with a 17 kJ/mol heat of adsorption).
Porous crystals simultaneously featuring void fractions and volumetric
surface areas in the 0.7–0.9 and 1300–1800 m2/cm3 ranges, respectively, were more susceptible to improvements
in deliverable capacity for the 100 bar/77 K ↔ 5 bar/160 K
swing by tuning their interactions with hydrogen. Select simulations
were analyzed in more detail to obtain adsorption mechanism insights.
Leveraging all of the generated data, we trained, for the first time,
a single artificial neural network capable of predicting hydrogen
loadings at multiple T, P conditions.
Using this neural network, we estimated that, for the nonisothermal
77 K ↔ 160 K swing, reducing the storage pressure from 100
to 35 bar only reduces the attainable deliverable capacity to 59 g/L,
which may be an acceptable trade-off due to safety and compression
cost implications. As the neural network only uses simple descriptors
as input, modelers and experimentalists alike could potentially use
it to rapidly pre-assess the hydrogen storage capabilities of newly
proposed crystal designs at various swing conditions.