Digital Hydrogen Platform (DigHyd): A Rigorously Curated Database for Hydrogen Storage Materials Empowered by AI-Assisted Literature Mining

The paper introduces DigHyd, an AI-assisted, rigorously curated database containing over 30,000 thermodynamic data entries for hydrogen storage materials that enables flexible evaluation of equilibrium behavior and supports data-driven discovery through validated composition-property relationships.

The Gist

Imagine you are trying to build the ultimate "fuel tank" for a hydrogen-powered car. You want it to be light, safe, and able to hold enough fuel to drive across the country. Scientists have been searching for the perfect material to do this for decades, testing thousands of different chemical recipes.

The problem? The information about these recipes is scattered across millions of research papers, written in different ways, and often missing the most important details. It's like trying to bake a cake using a library of recipes where some say "a pinch of salt," others say "10 grams," and some don't mention the oven temperature at all.

This paper introduces DigHyd (Digital Hydrogen Platform), a new, super-organized digital library designed to fix this mess. Here is how it works, explained simply:

1. The "AI Librarian" with a Human Check

Imagine a super-fast robot librarian (AI) that can read thousands of scientific papers in seconds. It pulls out numbers about how much hydrogen different materials can hold.

But robots can get confused. They might mix up units (like grams vs. atoms) or misread a graph. That's why DigHyd uses a "Human-in-the-Loop" approach. Think of the AI as a fast scanner that highlights the important pages, and a team of expert human scientists as the "editors" who double-check every single number. They ensure that the data is 100% accurate before it goes into the database.

2. Not Just "How Much," But "How Hard"

Most old databases just asked: "How much hydrogen can this material hold?" (This is like asking, "How big is the suitcase?").

DigHyd asks a much smarter question: "How much energy does it take to get the hydrogen in and out?" (This is like asking, "Is the suitcase heavy to lift, or does it have wheels?").

The scientists focused on two key "energy" numbers:

• Enthalpy (ΔH): How much heat is needed to release the hydrogen.
• Entropy (ΔS): How the disorder of the system changes.

By storing these "energy costs" instead of just a single pressure number, DigHyd is like a universal adapter. It allows engineers to calculate exactly how a material will behave in a hot desert or a freezing winter, rather than just giving them a number for one specific day.

3. The "Material Map"

The database organizes materials into different "neighborhoods":

• Interstitial Hydrides: The "veterans" of the group. They've been around a long time, are reliable, but don't hold a huge amount of fuel.
• Saline (Ionic) Hydrides: The "powerhouses." They hold a lot of fuel but are harder to work with (they need more heat to release it).

The database also shows how scientists have been "tweaking" these materials. It's like a recipe book that shows how adding a pinch of Nickel to a Magnesium base changes the taste (performance) of the final dish. This helps researchers see patterns: "Oh, whenever we add Element X, the material gets lighter but harder to use."

4. Testing the Crystal Ball (Machine Learning)

The team tested if this clean, organized data could help computers predict new materials. They used two types of "crystal balls":

• The "Black Box" (XGBoost): A powerful AI that guesses the answer based on patterns but doesn't explain why.
• The "White Box" (Symbolic Regression): A simpler AI that gives a clear mathematical formula, explaining the logic behind the guess.

The Result? Both crystal balls predicted the future performance of materials with almost the same high accuracy. This proves that the DigHyd database is so well-organized and consistent that even simple, logical rules can find the hidden patterns in the data.

Why This Matters

Before DigHyd, finding the right hydrogen storage material was like looking for a needle in a haystack while wearing blindfolds. Now, DigHyd is like a high-definition GPS for material scientists. It doesn't just list the needles; it tells you exactly where they are, how heavy they are, and how to get to them.

This platform gives researchers a solid foundation to design the next generation of hydrogen cars, ensuring they are safe, efficient, and ready for the real world.

Technical Summary

1. Problem Statement

Despite decades of research into solid-state hydrogen storage, data-driven discovery in this field has been hindered by two critical limitations in existing datasets:

• Inconsistent Reporting of Capacity: Gravimetric hydrogen storage density ($w$) is often reported inconsistently. Literature frequently conflates theoretical capacities (calculated from stoichiometry), reversible capacities (measured during cycling), and experimentally accessible capacities (under specific conditions), making cross-study comparison difficult.
• Lack of Systematic Thermodynamic Data: Fundamental thermodynamic parameters—specifically the enthalpy ($\Delta H$) and entropy ($\Delta S$) changes of hydrogenation reactions—are rarely reported systematically at a scale suitable for large databases. Most existing databases store equilibrium pressures at single temperatures, which limits the ability to evaluate material performance under diverse, application-specific operating conditions.
• Limitations of Pure AI Extraction: While AI and Large Language Models (LLMs) can extract vast amounts of data, they often lack the expert judgment required to ensure numerical correctness, particularly when converting units, distinguishing between different types of hydrogen content, or interpreting complex Pressure-Composition-Temperature (PCT) data.

2. Methodology

The authors developed DigHyd, a rigorously curated database constructed via an AI-assisted, human-in-the-loop framework. The workflow involves:

• AI-Assisted Mining: Scientific articles were identified via keyword searches. LLM-based tools extracted preliminary candidate data, including chemical composition, gravimetric density ($w$), PCT data, and thermodynamic quantities.
• Human-in-the-Loop Validation: AI-extracted data were treated as candidates only. Experts manually validated and curated the data to ensure accuracy.
  • Thermodynamic Derivation: For materials with multi-temperature PCT data, the team performed van't Hoff analysis to derive $\Delta H$ and $\Delta S$. The equilibrium plateau pressures ($P$) at various temperatures ($T$) were fitted to the equation:
    $$\ln P = -\frac{\Delta H}{RT} + \frac{\Delta S}{R}$$
    where $\Delta H$ is derived from the slope and $\Delta S$ from the intercept.
  • Cross-Checking: Directly reported $\Delta H$ and $\Delta S$ values were manually cross-checked against original PCT figures.
• Data Scope: The final curated dataset includes >4,000 experimental literature sources and >30,000 data entries. Specifically for metal hydrides, it contains 4,643 data points for $w$ and 652 data pairs for $\Delta H$ and $\Delta S$.
• Design Philosophy: Unlike databases storing fixed equilibrium pressures, DigHyd stores the fundamental thermodynamic parameters ($\Delta H, \Delta S$). This allows users to flexibly calculate equilibrium behavior for any specific operating temperature and pressure.

3. Key Contributions

• Creation of DigHyd: The first rigorously curated, thermodynamically grounded database for hydrogen storage materials, accessible at www.dighyd.org.
• Standardization of Thermodynamics: By prioritizing $\Delta H$ and $\Delta S$ over single-point equilibrium pressures, the database enables flexible, condition-specific performance evaluation.
• Hybrid Curation Framework: Demonstrated a successful workflow combining LLM-based extraction with expert human validation to overcome the "hallucination" and accuracy issues of pure AI mining.
• Compositional Mapping: The database not only catalogs material classes but also aggregates systematic modification strategies (alloying, doping) within representative systems (e.g., LaNi$_5$, MgH$_2$, LiBH$_4$).

4. Results and Analysis

• Temporal and Class Distribution: Analysis of the database reveals a steady increase in research output since the early stages of hydrogen storage. While interstitial hydrides remain the dominant class historically, there is a clear diversification toward complex hydrides, multi-component systems, and porous materials since the late 1990s.
• Thermodynamic Distinctions:
  • Gravimetric Density ($w$): Interstitial hydrides show a narrow distribution centered around 1.73 wt.%, whereas saline (ionic) hydrides exhibit substantially higher $w$ with a broader distribution.
  • Enthalpy ($\Delta H$): Saline hydrides have significantly higher $\Delta H$ values (more energetically demanding hydrogen release) compared to the moderate $\Delta H$ of interstitial hydrides.
  • Entropy ($\Delta S$): Distributions for $\Delta S$ partially overlap between classes, suggesting entropy is more influenced by gaseous H$_2$ contributions than specific bonding types.
• Machine Learning Validation:
  • The authors trained both physically interpretable symbolic regression models (white-box) and black-box XGBoost models to predict $w$ and room-temperature equilibrium pressure ($P_{eq,RT}$).
  • Result: Both models achieved comparable predictive performance ($R^2$, RMSE, MAE). This confirms that the curated dataset possesses internal consistency and contains learnable, physically meaningful composition-property relationships.
  • The symbolic regression models provided explicit analytical expressions, validating that the data supports physically transparent modeling.

5. Significance

• Foundation for Data-Driven Discovery: DigHyd provides a reliable, high-quality dataset that overcomes the noise and inconsistency of raw literature, enabling robust statistical analysis and machine learning.
• Engineering Relevance: By storing fundamental thermodynamic parameters, the platform allows researchers to simulate material performance under specific real-world operating conditions (e.g., specific temperature/pressure windows) rather than relying on static, single-point data.
• Scalability and Future Outlook: The framework is designed to be extensible. Future iterations will include porous materials, complex hydrides, kinetic descriptors, and the development of "physically constrained AI Agents" for autonomous materials discovery.
• Paradigm Shift: The paper advocates for a shift from superficial data collection to thermodynamically rigorous curation, ensuring that AI-driven materials science is built on physically correct foundations.