2026-10-23
Tohoku's DigHyd mines papers with LLMs, then humans recompute enthalpy and entropy from PCT curves. Capacity test R² is 0.868; pressure error still exceeds 10×.
Whether a solid stores hydrogen usefully depends on more than how much hydrogen fits in the solid. It also depends on the equilibrium pressure at the temperature you actually run. The literature does not put those numbers in one table. Gravimetric density is sometimes a theoretical value read off the formula, sometimes the reversible capacity over cycles, sometimes a measurement at one pressure and temperature. A few papers even report hydrogen atoms per metal atom and leave it looking like a weight percent. The enthalpy and entropy that set equilibrium pressure sit inside pressure-composition-temperature curves and almost never arrive as database fields. Language models can now copy reported numbers at scale. They do not fix unit mismatches, and they do not replace a person reading plateau pressures off a multi-temperature plot.
DigHyd comes from Hao Li and Shin-ichi Orimo's group at Tohoku University's WPI-AIMR, with collaborators at the University of Tokyo. Construction has two stages.
Keyword and bibliographic searches collect papers on hydrogen storage and metal hydrides. An LLM pulls candidate composition, reported gravimetric density, PCT data, and any thermodynamic values the text already states, and it keeps the DOI. Those extractions stay candidates.
The thermodynamic table is redone by hand. For metal hydrides, enthalpy and entropy refer to the hydrogenation reaction of metal plus hydrogen gas. When a paper shows PCT curves at several temperatures, a curator reads the plateau pressure at each temperature and runs a van't Hoff plot: the log of equilibrium pressure against inverse temperature, slope for enthalpy, intercept for entropy. Values printed in the paper are checked against the figure when the figure exists. For metal hydrides alone, that pass produced 4,643 gravimetric-density rows and 652 enthalpy-entropy pairs. The full platform holds more than 4,000 experimental papers and more than 30,000 entries. The thermodynamic layer is the narrow part.
The database stores enthalpy and entropy, not equilibrium pressure at one frozen temperature. Change the operating temperature and the pressure can be recomputed. Room-temperature equilibrium pressure is that derived number. It folds both the heat of reaction and the entropy of hydrogen gas into one engineering target.
Interstitial hydrides have a tight gravimetric-density distribution centered near 1.73 wt%, over 2,859 rows. Saline ionic hydrides, 1,785 rows, sit higher and wider, and their enthalpies are larger, so hydrogen release costs more energy. Entropy distributions overlap. Gaseous hydrogen dominates the entropy, so material class does not separate it cleanly. The thermodynamic sample is smaller: 371 interstitial pairs and 281 saline pairs, which sum to the 652. Inside LaNi5, MgH2, and LiBH4 families, elemental substitutions and additives spread capacity, equilibrium pressure, and absorption temperature about as widely as the gap between classes.
The regression uses an 80:20 train-test split and compares symbolic regression with XGBoost. The capacity axis in the figure is log10(w/⟨M⟩), a log capacity normalized by average atomic mass, not raw weight percent.
| Model | Target | Test R² | Test RMSE | Test n |
| Symbolic regression | Capacity | 0.800 | 0.220 | 394 |
| XGBoost | Capacity | 0.868 | 0.164 | 394 |
| Symbolic regression | Room-temp. pressure | 0.750 | 1.418 log10(MPa) | 216 |
| XGBoost | Room-temp. pressure | 0.786 | 1.281 log10(MPa) | 216 |
XGBoost wins on capacity. On 1,573 training rows its R² is 0.930, against 0.828 for the symbolic model. On pressure, both test errors exceed 1 log10(MPa), more than a factor of ten in pressure. XGBoost's training R² of 0.992 and RMSE of 0.248 fall to a test RMSE of 1.281. The prose calls the two models comparable.
Screening needs thermodynamics you can reevaluate at your own temperature, not a pressure frozen at 25°C or 300°C in one paper. The 652 pairs were checked against PCT plots. That is the difference between this table and a raw LLM extract. Symbolic regression leaves an explicit formula in elemental features. XGBoost fits capacity more tightly. Composition and storage behavior in this table are learnable. The article is a commentary in Applied Physics A. The class distributions and part of the curves are carried over from earlier Chemical Science papers. The new comparison is XGBoost. Treat it as a curated data floor plus one regression study.
Thermodynamic coverage is much thinner than the capacity table. Strict enthalpy and entropy for porous materials and most complex hydrides are listed as future work, so the class comparison is mainly interstitial versus saline hydrides. A test RMSE above one order of magnitude in pressure will mis-rank materials against a 1–10 bar window. The train-test gap on XGBoost pressure is larger than the word "comparable" suggests. The regressions use 1,967 capacity rows and 1,078 pressure rows, and the prose does not state the filter relative to 4,643 and 652. Extraction error rates for the LLM stage are not reported, so the amount of human correction stays invisible.