Eight numbers, added together. That is the whole model.
Ali Ebrahimpoor Gorji and Ville Alopaeus, chemical engineers at Aalto University in Finland, spent their paper arguing that predicting how much hydrogen dissolves in a liquid does not require a neural network, a random forest, or anything else you cannot write on an index card. Their final equation multiplies the logarithm of temperature by 1.81, the logarithm of pressure by 1.00, adds six numbers describing the molecule's structure, subtracts 20.95, and stops.
The question matters because hydrogen keeps being proposed as a fuel, and fuels have to be stored, moved, and reacted in liquids. How much hydrogen a given solvent will hold, at a given temperature and pressure, shapes the design of storage systems and of the refinery processes that add hydrogen to oil fractions. Measuring it in the lab is slow. Predicting it from an equation is not.
The authors assembled measurements from 62 published studies, covering 101 chemicals across eleven families: alkanes, alcohols, aromatics, ethers, esters, ketones, halides and more. Then they threw a lot of it away. Fifty-nine duplicate readings, three that contradicted each other, 476 taken at extreme pressures, eight at extreme temperatures, and 30 others went into an exclusion list they published alongside the paper. What survived was 3,761 measurements spanning temperatures from 88 to 623 kelvin and pressures from 1 to 299 bar, plus eight of their own earlier measurements of hydrogen in octadecane.
Choosing what to feed the model
Most previous work described each solvent using its critical temperature, critical pressure and molecular weight: standard thermodynamic properties, tabulated for common compounds. Gorji and Alopaeus think that is the wrong vocabulary. Critical properties describe how a substance behaves near its critical point; they say surprisingly little about whether a molecule has a hydroxyl group hanging off it.
So the team drew each of the 100 molecules by hand in ChemBioDraw and fed the structures to PaDEL, free software that computes numerical descriptors of molecular shape and composition. After discarding descriptors that barely varied, they had a pool of 1,239 candidates. A genetic algorithm, which breeds and mutates candidate variable sets the way evolution breeds organisms, picked the six that mattered. Among them: a count of single bonds, a measure of mean ionization potential, and SsOH, which quantifies hydroxyl groups specifically. That last one is the tell. The authors' own earlier model, trained only on hydrocarbons, had no way to represent an -OH group and stumbled on alcohols.
The comparison they run on themselves is the convincing part. Fed the same 2,570 training points, a model using only temperature and pressure managed an R-squared of 0.65 on the log-transformed data. Swap in the conventional critical properties and it reaches 0.95. Swap in the PaDEL structural descriptors instead and it reaches 0.96, with a leave-one-out cross-validated score of 0.96 as well, meaning the fit does not collapse when you hide individual points from it.
On 1,191 held-out measurements from 24 chemicals the model had never seen, R-squared was 0.93 on log values. Critically, that is not much below the training figure. The most recent competing study, which used decision trees and neural networks on a nearly identical dataset, reported a perfect 1.000 on its training set and 0.934 on its test set, a gap the Aalto authors read as memorization rather than learning.
Four of the chemicals in the test set are ones the earlier study had to discard entirely, because nobody has measured their critical properties. Molecular descriptors can be computed from a drawing, so those four went straight through.
Why it matters
The honest number is not the R-squared. Converted back from logarithms to actual solubility, the model's average absolute relative deviation is 19 percent on the training set and 23 percent on the test set. That is a real error bar, and the paper reports it plainly. Against the black-box models, the linear equation beat all four on some test chemicals and lost on others; acetonitrile, in particular, it handles poorly.
What the equation buys instead is transparency and reach. You can see which structural features push hydrogen solubility up or down, and you can apply it to any molecule you can sketch, including compounds no one has characterized thermodynamically. Screening a hundred candidate solvents for a hydrogen storage system is a different exercise when the barrier is drawing structures rather than finding critical-point data.
The authors are careful about the limits. Three outliers in the dataset all trace to carbon disulfide, the only sulfur-containing compound present, and the model's applicability domain, the region of chemical space where its predictions can be trusted, is bounded by what the training set covers. They ask for more experimental measurements on underrepresented chemicals. The claim they actually make is narrow and worth taking seriously: for this property, the extra machinery was not paying for itself.