UniHam: The First SOC-Complete Large-Scale Hamiltonian Dataset and Benchmark

UniHam releases a 100,000+ DFT Hamiltonian dataset with SOC and a three-pillar benchmark, revealing model failure modes.
UniHam is a newly proposed Hamiltonian learning dataset and benchmark suite containing over 100,000 DFT-computed complex-valued Hermitian Hamiltonians, covering 72 elements with full spin-orbit coupling (SOC). The benchmark evaluates models on Hamiltonian reconstruction accuracy, out-of-distribution (OOD) generalization, and downstream property prediction. Experiments reveal systematic failure modes in SOC- and element-related scenarios and significant differences in composition OOD robustness, emphasizing the need for spectral-level evaluation. UniHam offers a reproducible, SOC-complete standard for quantum materials foundation models.
Why Hamiltonian Learning Needs a New Benchmark
Accurately predicting electronic Hamiltonians allows material property inference to bypass the high computational cost of density functional theory (DFT)—this is the core vision of foundation models for quantum materials. But a data bottleneck has persisted: existing Hamiltonian datasets are small, lack structural diversity, and often omit critical relativistic physics effects such as spin-orbit coupling (SOC). A new paper proposes UniHam to address exactly this gap—it builds a dataset and benchmark suite containing 100,000+ complex-valued Hermitian Hamiltonians obtained from DFT calculations, covering 72 elements with full SOC.

The electronic Hamiltonian in quantum mechanics describes the total energy operator of a system; its matrix form (under a given basis set) contains all information including electron kinetic energy, potential energy, and electron-electron interactions. For crystalline materials, the eigenvalues and eigenvectors of the Hamiltonian directly determine electronic properties such as band structure and density of states. Therefore, if machine learning models can accurately predict Hamiltonians, they can bypass expensive DFT self-consistent calculations and rapidly infer material properties. However, Hamiltonian matrices are typically Hermitian (real symmetric or complex Hermitian), and their dimensionality grows sharply with basis set size—learning their high-dimensional complex-valued structure places high demands on both model capacity and data scale.
The UniHam Dataset: Scale and Physical Completeness
UniHam's construction logic revolves around two words: scale and completeness. All 100,000+ Hamiltonians are obtained from DFT calculations and are complex-valued Hermitian matrices, guaranteeing physical validity. Covering 72 elements means the dataset spans a considerably broad chemical space; in terms of crystal geometry and symmetry, it encompasses multiple lattice types and space group families, providing a foundation in structural diversity for both training and evaluation.
The inclusion of SOC is what sets this dataset apart from prior work. Spin-orbit coupling cannot be neglected in heavy-element systems; omitting it distorts Hamiltonian predictions when electronic structure details are involved. UniHam makes SOC a default configuration rather than an optional one, which directly determines whether it can support general-purpose models for real materials.
Spin-orbit coupling (SOC) is a relativistic effect arising from the interaction between electron spin and orbital motion. In heavy elements (such as transition metals and lanthanides), SOC is significant and leads to physical phenomena such as band splitting and topologically nontrivial states. If SOC is neglected in DFT calculations, the resulting Hamiltonian lacks these critical details, which in turn affects the accurate description of a material's electronic structure. UniHam making SOC a default configuration means its Hamiltonian matrices are complex-valued Hermitian matrices that more faithfully reflect the physical behavior of systems containing heavy elements, providing the physical completeness necessary for training general-purpose materials models.
The Three-Pillar Evaluation Protocol
On top of UniHam, the authors designed a standardized protocol to evaluate two representative frontier models and proposed three complementary evaluation dimensions.
Hamiltonian reconstruction accuracy: This is the most direct metric, measuring how close the model-predicted matrix is to the DFT reference.
Out-of-distribution (OOD) generalization: This evaluates model performance under shifts in composition and symmetry. The paper specifically notes that different models exhibit significant differences in composition OOD robustness—meaning that looking only at in-distribution accuracy can mask risks in real-world deployment.
Downstream property prediction capability: Matrix accuracy alone is insufficient; the key question is whether the predicted Hamiltonian can reliably support electronic structure property inference. To this end, the authors emphasize the necessity of spectral-level evaluation—only validation at the energy spectrum level can determine whether Hamiltonian predictions are truly usable.
Out-of-distribution (OOD) generalization refers to model performance on data outside the training distribution. In materials science, training and test sets may exhibit distribution shifts due to factors such as elemental composition and crystal symmetry. For example, a model may perform significantly worse when predicting Hamiltonians for element combinations or space groups it has not seen during training. OOD generalization evaluation aims to expose this risk, ensuring that models not only perform well on the in-distribution set but can also be reliably deployed in real-world scenarios. UniHam systematically evaluates model robustness through composition and symmetry shifts, addressing the shortcoming of relying solely on in-distribution accuracy.
Failure Modes Revealed by the Benchmark
Experiments on UniHam yielded several noteworthy findings. The performance of the two representative models was effectively differentiated by this metric suite, demonstrating that the benchmark has genuine discriminative power rather than being saturated. More specifically, the experiments exposed intrinsic failure modes related to SOC and elements—models systematically mispredict in certain element or SOC-related scenarios, and such issues tend to be averaged out in aggregate metrics. The large fluctuations in composition OOD robustness also confirm that generalization evaluation cannot be omitted. The necessity of spectral-level evaluation conversely implies that matrix-level error alone may overestimate a model's actual value for downstream tasks.
These conclusions have direct reference value for researchers working on materials foundation models: the design of the evaluation protocol itself influences your judgment of model capability.
Significance for Quantum Materials Foundation Models
UniHam is positioned as a reproducible, SOC-complete benchmark to sharpen comparisons between models and accelerate the development of next-generation quantum materials foundation models. The combination of dataset, benchmark, and accompanying metrics gives different works a common basis for comparison. For the field, this addresses not just the issue of data volume but also the lack of evaluation standards—progress becomes measurable when everyone compares using the same protocol and the same samples covering SOC and space group diversity.
The paper also open-sources the dataset and benchmark (see the original paper for access details), offering a low-barrier starting point for research teams wishing to reproduce or extend this work.
One-line Summary
UniHam releases a 100,000+ DFT Hamiltonian dataset with SOC and a three-pillar benchmark, revealing model failure modes.
Paragraph Summary
UniHam is a newly proposed Hamiltonian learning dataset and benchmark suite containing over 100,000 DFT-computed complex-valued Hermitian Hamiltonians, covering 72 elements with full inclusion of spin-orbit coupling (SOC). The benchmark evaluates models along three dimensions: Hamiltonian reconstruction accuracy, out-of-distribution (OOD) generalization (composition and symmetry shifts), and downstream property prediction. Experiments reveal systematic failure modes of two representative models in SOC- and element-related scenarios, as well as significant differences in composition OOD robustness, and emphasize the necessity of spectral-level evaluation—matrix-level error alone may overestimate a model's actual value for downstream tasks. UniHam provides a reproducible, SOC-complete evaluation standard for quantum materials foundation models, helping to advance progress in the field.
Related articles

Zed Editor Releases v1.29.2: Observations on a Routine Version Iteration
Zed editor releases patch v1.29.2, version bump by @JosephTLyons, commit verified via GitHub signature. This article covers the release's basic info, what a patch version means, and practical upgrade advice for users.

Anthropic IPO Questioned: Hacker News Debates Timing of Going Public
A Hacker News post titled "Just say no to Anthropic's IPO" sparks debate on the timing of Anthropic's public listing and AI capitalization, touching on valuation bubbles and mission independence.

Hands-On with the Genie 4 World Model: Directing Worlds in Real Time Like a Puppeteer
Google DeepMind's Genie 4 world model in depth: Special Guest X's hands-on demo of real-time scene direction via prompts, custom interactive apps in Project Genie, and prompts-as-seeds reproducibility.