[论文解读] A generative machine learning model for designing metal hydrides applied to hydrogen storage

Xiyuan Liu, Christian Hacker+2 more|arXiv (Cornell University)|Jan 28, 2026

Hydrogen Storage and Materials被引用 0

一句话总结

论文提出一种因果发现引导的轻量化生成式 ML 框架（CDVAE + FCI）来设计用于氢储存的新金属氢化物，生成 1,000 个候选材料并识别出六种未报道的化合物，其中四种通过了 DFT 验证。

ABSTRACT

Developing new metal hydrides is a critical step toward efficient hydrogen storage in carbon-neutral energy systems. However, existing materials databases, such as the Materials Project, contain a limited number of well-characterized hydrides, which constrains the discovery of optimal candidates. This work presents a framework that integrates causal discovery with a lightweight generative machine learning model to generate novel metal hydride candidates that may not exist in current databases. Using a dataset of 450 samples (270 training, 90 validation, and 90 testing), the model generates 1,000 candidates. After ranking and filtering, six previously unreported chemical formulas and crystal structures are identified, four of which are validated by density functional theory simulations and show strong potential for future experimental investigation. Overall, the proposed framework provides a scalable and time-efficient approach for expanding hydrogen storage datasets and accelerating materials discovery.

研究动机与目标

• Identify key features causally related to hydrogen storage performance to reduce data requirements.
• Develop a lightweight generative model that can propose novel metal hydrides from small datasets.
• Integrate causal discovery with a generative model to produce crystal-structure–aware candidates.
• Validate generated candidates using DFT and a fast relaxation model to guide experimental exploration.

提出的方法

• Define a Hydrogen Storage Score combining hydrogen weight fraction and formation energy with a modified energy factor.
• Apply Fast Causal Inference (FCI) to identify the causal neighborhood of the storage score and select features.
• Train a Crystal Diffusion Variational Autoencoder (CDVAE) on 450 MP-based observations to generate new formulas and CIFs.
• Relax generated structures with M3GNet to obtain feasible crystal structures and recalculate properties.
• Use DFT (VASP) to compute formation energies for validation and filter top candidates.
• Filter and rank generated candidates to identify top materials for experimental consideration.

实验结果

研究问题

• RQ1 Can causal discovery identify a minimal, causally relevant feature subset predictive of hydrogen storage performance in metal hydrides?
• RQ2 Can a lightweight CDVAE-based generator produce chemically valid, novel metal hydrides not in existing databases?
• RQ3 Do DFT-validatable candidates emerge from the generative process, and how many are practically feasible for hydrogen storage applications?

主要发现

Formula	E_form Predicted (eV/atom)	E_form DFT (eV/atom)	H Storage Score	Squared error	MP Unique ID	Same Formula	Same Ratio
Li 3 B 3 H 6	0.057	-0.189	0.062	0.061	mp-568523	FALSE	TRUE
Li 1 Al 3 H 6	0.019	0.049	0.043	0.001	FALSE	FALSE	-
Ti 1 H 2	-0.426	-0.466	0.040	0.002	mp-1077482	TRUE	TRUE
Ti 2 H 4	-0.401	-0.465	0.040	0.004	mp-1077482	FALSE	TRUE
K 2 Al 3 H 6	-0.159	0.080	0.032	0.057	FALSE	FALSE	-
Ti 6 H 8	-0.387	-0.315	0.027	0.005	FALSE	FALSE	-
Ti 3 H 4	-0.322	-0.171	0.026	0.023	FALSE	FALSE	-
Ti 5 H 6	-0.361	-0.350	0.024	0.000	FALSE	FALSE	-
Ca 2 Al 1 Si 2 H 3	-0.404	-0.243	0.018	0.026	FALSE	FALSE	-
Ti 4 Ni 1 H 4	-0.390	-0.372	0.016	0.000	FALSE	FALSE	-

• FCI 分析识别出氢重量分数、形成能和晶体结构是氢储存分数的关键预测因子。
• 在 450 个训练样本中，CDVAE 生成了 1,000 个候选材料，识别出六种未报道的合金氢化物，四种通过了 DFT 筛选。
• 整合的 HDL 流程在与 M3GNet-J 相关的放松下实现了与 DFT 兼容的预测，形成能的 MAE 约为 0.0775 eV，总 MSE 约为 0.018（在生成集合上）。
• 四个经过验证的候选在计算测试中表现出有利的氢储存特性，表明具有较强的实验后续潜力。