# arXiv cs.AI: Physical structure as an organizing principle for universal Hamiltonian learning [Canonical reading](https://escapemeridian.com/readings/arxiv-cs-ai-physical-structure-as-an-organizing-principle-for-0995a7c4) The authors introduce UniTB, a machine-learned Hamiltonian designed to work across different material compositions while keeping calculated properties physically plausible. It combines physical priors from density-functional tight-binding theory with deep learning; a mixture of experts and a charge-equilibration head help it generalize across chemistry. The authors say a projection limits band shifts from the prior, suppressing spurious states. They also present UniTB-Flash, which assembles its prior from two-center tables at O(N) cost for N atoms, compared with O(N log N) on a grid. Trained on 29,242 crystals, UniTB reaches a reported 0.48 meV matrix-element error on nonmagnetic test crystals spanning 68 elements. The authors also report agreement with density functional theory (DFT) for optical and conductivity properties of amorphous Ge-Sb-Te, despite having no amorphous training data. In a screening of 16,195 crystals across five application domains, UniTB-Flash ranks candidates with no significant difference from UniTB, while UniTB's property values are more accurate. The abstract reports these results but does not establish how well either model performs against experimental measurements or in practical materials screening. Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims. ## Original source arXiv cs.AI · Article Original source date: Oct 26, 2026 (2026-10-26). Date basis: Date basis not provided. Added to Escape Meridian: 2026-11-13T22:17:23.180Z. First observed: 2026-10-28T07:36:34.315Z. Retrieved: 2026-10-29T17:03:10.641Z. - [Physical structure as an organizing principle for universal Hamiltonian learning](https://arxiv.org/abs/2610.36135) ## Limits The evidence is an arXiv abstract; it does not provide details of the test-set construction, the comparisons, or validation against experimental measurements. The abstract reports a screening of 16,195 crystals across five application domains but does not identify the domains or establish practical screening outcomes. ## Open question How do UniTB and UniTB-Flash compare with experimental measurements when used to screen materials beyond the reported test sets? ## Source-stated The authors present UniTB as a universal machine-learned Hamiltonian framework that combines physical priors from density-functional tight-binding theory with deep learning. > Here we introduce UniTB, a universal E(3) equivariant Hamiltonian representation framework that integrates physical priors rooted in density-functional tight-binding theory with deep learning. - [Supporting source](https://arxiv.org/abs/2610.36135) ## Source-stated The authors say UniTB is designed to generalize across chemistry and constrain band shifts to suppress spurious states. > generalizes across chemistry through a mixture of experts routed by that prior and a charge-equilibration head - [Supporting source](https://arxiv.org/abs/2610.36135) ## Source-stated The authors report that UniTB-Flash assembles its prior from two-center tables at O(N) cost, compared with O(N log N) on a grid. > UniTB-Flash assembles its prior from two-center tables at $O(N)$ cost for $N$ atoms, against $O(N\\log N)$ on a grid. - [Supporting source](https://arxiv.org/abs/2610.36135) ## Source-stated On the reported test crystals, UniTB had a 0.48 meV matrix-element error; the test set covered nonmagnetic crystals spanning 68 elements. > Trained on 29,242 crystals, UniTB reaches a 0.48 meV matrix-element error on nonmagnetic test crystals spanning 68 elements. - [Supporting source](https://arxiv.org/abs/2610.36135) ## Source-stated The authors report that both models agree with DFT optics and conductivity for amorphous Ge-Sb-Te despite having no amorphous training data. > Without amorphous training data, both models agree with DFT optics and conductivity of amorphous Ge-Sb-Te. - [Supporting source](https://arxiv.org/abs/2610.36135) ## Source-stated For the reported screening, UniTB-Flash ranked candidates with no significant difference from UniTB, whose property values were more accurate. > UniTB-Flash ranks candidates with no significant difference from UniTB, whose property values are more accurate; together they screen 16,195 crystals for five application domains of industrial relevance. - [Supporting source](https://arxiv.org/abs/2610.36135) ## Source and contributor context Liu, Mingkang, Zhouyin, Zhanghao, Wang, Guanghui, Zheng, Daye, Shen, Lei, Gu, Qiangqiang · Article author These are credits stated by the source; they do not establish a profile match or authorship of every passage. ## Source media No suitable media was found by the supported page extraction. ## Terms and references Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement. ## Deep learning Machine learning that uses many layers of simple calculations; the basis of most current AI. - [Definition reference](https://en.wikipedia.org/wiki/Deep_learning) ## Mixture of experts A model design that sends each input to a few specialized sub-networks instead of the whole model, which saves computation. - [Definition reference](https://developers.google.com/machine-learning/glossary#mixture-of-experts) ## Held-out set Data kept aside during training and used only to test how well a model works on examples it has not seen. - [Definition reference](https://developers.google.com/machine-learning/glossary#holdout-data) ## arXiv arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed. - [Explore this context](/entities/arxiv) - [Reference](https://arxiv.org/) - [Reference](https://info.arxiv.org/help/api/index.html) ## Editorial provenance Version 1. Publication ID: 526e4d35-5684-42c8-8963-6848506c1c93. Prepared by gpt-6-luna; grounding review by gpt-6-luna. Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-13T22:17:23.180Z. This checks fidelity to the source, not whether its claims are independently true. Editorial selection: accept; reasons: material\_development. Observation: d9e5b6bd37612be3dc8d56193f845392f401db10465f85d0786d326fac11a140. Source text hash: 61bb8f2726dceaca2c3aea1222954bf3837a90c41ee14ea63a7d9024780d4a56. ## Revision and source-check history ### Version 1 — 2026-11-13T22:17:23.180Z The authors introduce UniTB, a machine-learned Hamiltonian designed to work across different material compositions while keeping calculated properties physically plausible. It combines physical priors from density-functional tight-binding theory with deep learning; a mixture of experts and a charge-equilibration head help it generalize across chemistry. The authors say a projection limits band shifts from the prior, suppressing spurious states. They also present UniTB-Flash, which assembles its prior from two-center tables at O(N) cost for N atoms, compared with O(N log N) on a grid. Trained on 29,242 crystals, UniTB reaches a reported 0.48 meV matrix-element error on nonmagnetic test crystals spanning 68 elements. The authors also report agreement with density functional theory (DFT) for optical and conductivity properties of amorphous Ge-Sb-Te, despite having no amorphous training data. In a screening of 16,195 crystals across five application domains, UniTB-Flash ranks candidates with no significant difference from UniTB, while UniTB's property values are more accurate. The abstract reports these results but does not establish how well either model performs against experimental measurements or in practical materials screening. ## Original source arXiv cs.AI · Article Original source date: Oct 26, 2026 (2026-10-26). Date basis: Date basis not provided. Added to Escape Meridian: 2026-11-13T22:17:23.180Z. First observed: 2026-10-28T07:36:34.315Z. Retrieved: 2026-10-29T17:03:10.641Z. - [Physical structure as an organizing principle for universal Hamiltonian learning](https://arxiv.org/abs/2610.36135) ## Limits The evidence is an arXiv abstract; it does not provide details of the test-set construction, the comparisons, or validation against experimental measurements. The abstract reports a screening of 16,195 crystals across five application domains but does not identify the domains or establish practical screening outcomes. ## Open question How do UniTB and UniTB-Flash compare with experimental measurements when used to screen materials beyond the reported test sets? ## Source-stated The authors present UniTB as a universal machine-learned Hamiltonian framework that combines physical priors from density-functional tight-binding theory with deep learning. > Here we introduce UniTB, a universal E(3) equivariant Hamiltonian representation framework that integrates physical priors rooted in density-functional tight-binding theory with deep learning. - [Supporting source](https://arxiv.org/abs/2610.36135) ## Source-stated The authors say UniTB is designed to generalize across chemistry and constrain band shifts to suppress spurious states. > generalizes across chemistry through a mixture of experts routed by that prior and a charge-equilibration head - [Supporting source](https://arxiv.org/abs/2610.36135) ## Source-stated The authors report that UniTB-Flash assembles its prior from two-center tables at O(N) cost, compared with O(N log N) on a grid. > UniTB-Flash assembles its prior from two-center tables at $O(N)$ cost for $N$ atoms, against $O(N\\log N)$ on a grid. - [Supporting source](https://arxiv.org/abs/2610.36135) ## Source-stated On the reported test crystals, UniTB had a 0.48 meV matrix-element error; the test set covered nonmagnetic crystals spanning 68 elements. > Trained on 29,242 crystals, UniTB reaches a 0.48 meV matrix-element error on nonmagnetic test crystals spanning 68 elements. - [Supporting source](https://arxiv.org/abs/2610.36135) ## Source-stated The authors report that both models agree with DFT optics and conductivity for amorphous Ge-Sb-Te despite having no amorphous training data. > Without amorphous training data, both models agree with DFT optics and conductivity of amorphous Ge-Sb-Te. - [Supporting source](https://arxiv.org/abs/2610.36135) ## Source-stated For the reported screening, UniTB-Flash ranked candidates with no significant difference from UniTB, whose property values were more accurate. > UniTB-Flash ranks candidates with no significant difference from UniTB, whose property values are more accurate; together they screen 16,195 crystals for five application domains of industrial relevance. - [Supporting source](https://arxiv.org/abs/2610.36135) ## Source and contributor context Liu, Mingkang, Zhouyin, Zhanghao, Wang, Guanghui, Zheng, Daye, Shen, Lei, Gu, Qiangqiang · Article author These are credits stated by the source; they do not establish a profile match or authorship of every passage. ## Source media No suitable media was found by the supported page extraction. ## Terms and references Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement. ## Deep learning Machine learning that uses many layers of simple calculations; the basis of most current AI. - [Definition reference](https://en.wikipedia.org/wiki/Deep_learning) ## Mixture of experts A model design that sends each input to a few specialized sub-networks instead of the whole model, which saves computation. - [Definition reference](https://developers.google.com/machine-learning/glossary#mixture-of-experts) ## Held-out set Data kept aside during training and used only to test how well a model works on examples it has not seen. - [Definition reference](https://developers.google.com/machine-learning/glossary#holdout-data) ## arXiv arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed. - [Explore this context](/entities/arxiv) - [Reference](https://arxiv.org/) - [Reference](https://info.arxiv.org/help/api/index.html) ## Editorial provenance Version 1. Publication ID: 526e4d35-5684-42c8-8963-6848506c1c93. Prepared by gpt-6-luna; grounding review by gpt-6-luna. Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-13T22:17:23.180Z. This checks fidelity to the source, not whether its claims are independently true. Editorial selection: accept; reasons: material\_development. Observation: d9e5b6bd37612be3dc8d56193f845392f401db10465f85d0786d326fac11a140. Source text hash: 61bb8f2726dceaca2c3aea1222954bf3837a90c41ee14ea63a7d9024780d4a56. ### uncertain — 2026-11-13T22:17:23.180Z — anthropic/claude-opus-5.5 Publication: 526e4d35-5684-42c8-8963-6848506c1c93 attribution — pass: The reading attributes UniTB, UniTB-Flash and the results to the paper's authors. It labels results as reported rather than as established facts. The arXiv listing and author list support this. > Authors:Mingkang Liu, Zhanghao Zhouyin, Guanghui Wang, Daye Zheng, Lei Shen, Qiangqiang Gu summary — uncertain: The reading is faithful but hard for non-specialists to follow. It drops the abstract's motivation: DFT is costly, which limits large calculations and materials screening, and system-specific models do not generalize. It also never says plainly what a machine-learned Hamiltonian does. DFT is spelled out only in the second paragraph. Central terms (Hamiltonian, tight-binding, matrix-element error, meV, band shifts, spurious states, O(N)) need plain-language glosses or tooltips. Smallest repair: open with the cost problem and a one-line explanation of the Hamiltonian's role. > Density functional theory (DFT) gives the electronic structure of materials, but its cost limits large-supercell calculations and materials screening. date — pass: originalPublished is 2026-10-26, which matches the v1 submission date. The later feed date of 2026-10-29 is correctly not used. > \[Submitted on 26 Oct 2026\] support — pass: Every numeric and technical claim matches the abstract: 29,242 crystals, 0.48 meV error, 68 elements, O(N) versus O(N log N), amorphous Ge-Sb-Te agreement, and 16,195 crystals screened across five domains. The band-shift projection claim is also supported, though claim 2's excerpt covers only the generalization part. The limitations and the question about experimental validation stay appropriately uncertain and add no unsupported benefits. > A projection bounds band shifts from the prior, suppressing spurious states and keeping downstream properties physical. media — pass: The selected media is the verified direct link to the original arXiv abstract page, which is the actual source. ## Recorded references Explicit citations in public readings and public file attachments to this reading. Shared topics are not citations. Other artifact types need an explicit reference adapter. No public references recorded here yet. [Reference data (JSON)](/readings/0995a7c45bd05c00eadd4743ad25dd9d600301f8957cd1f2411d30dcf924980d/references.json) ## Corrections and challenges [Send a correction or an opposing source](https://escapemeridian.com/corrections/) [JSON record](https://escapemeridian.com/readings/arxiv-cs-ai-physical-structure-as-an-organizing-principle-for-0995a7c4.json)