Machine-learning models that predict whether a material superconducts, and at what temperature, mostly learn from the positions of atoms in real space. Physics points elsewhere for an important class of these materials. Many unconventional superconductors and charge-ordered metals sit close to a Van Hove singularity: a saddle point in the electron energy landscape where the number of available electron states spikes. When the material's filling level lands near that spike, electrons crowd together and the metal becomes unstable. We propose to describe each material with a handful of numbers about those saddle points. The descriptors are computed from standard band-structure calculations and shifted by each compound's doping level. We will test whether a deliberately small model built on them predicts superconductivity in chemical families it has never seen better than large real-space neural networks do.
The machine-learning half of the question is the familiar tradeoff between model complexity and generalization. Large models fit their training data better, and on random train/test splits they look better. They may also be memorizing family-specific chemical proxies. A cuprate "looks like" a cuprate whether or not the model understands why it superconducts. Our claim has two parts. First, a low-capacity hypothesis class whose inputs encode the actual instability mechanism should win when whole families are held out. Second, standard random-split cross-validation will choose the wrong model. If both hold, materials benchmarks are rewarding the wrong thing. If neither holds, that is a clean negative result about how far rigid-band saddle-point physics transfers across materials.
Core Claim
We will compute saddle-point descriptors from first-principles band structures and place the filling level by a rigid-band shift for each doped compound. With a few tolerance-aware features about the nearest saddle point, a sparse linear or shallow tree model should match or beat a pretrained crystal-graph network on leave-one-family-out superconductor prediction, while trailing it on random splits. The ingredient we expect to matter most is the saddle-point information itself. A generic density of states at the filling level should not be enough.
Grounding in Prior Work
The physics literature motivates the descriptors and also limits how they should be built. Work on the honeycomb Hubbard model (Van Hove tuning of Fermi surface instabilities through compensated metallicity) uses diagrammatic calculations to show that moving a hole pocket near a saddle point selects a chiral d-wave superconductor. Studies of kagome metals (Analysis of Charge Order in the Kagome Metal AV3Sb5 (A=K,Rb,Cs); Discovery of charge order above room-temperature in the prototypical kagome superconductor La(Ru1−xFex)3Si2) tie charge order to filling near a saddle point. Work on SrPtAs (Chiral d-wave superconductivity in SrPtAs) ties chiral pairing to a high density of states at the saddle. Together these papers say what to measure. All of them study one material or one model at a time.
Two results constrain the design. Self-consistent Hartree-Fock calculations on twisted bilayer graphene (Does filling-dependent band renormalization aid pairing in twisted bilayer graphene?) show that electron-electron interactions pull the saddle point onto the filling level across a whole range of fillings. Correlated quasiparticle calculations on monolayer NbSe2 (Quasiparticle energy bands and Fermi surfaces of monolayer NbSe2) show that correlations reshape where the saddle points sit. The consequence is that the exact energy distance between a saddle point and the filling level, as computed by a non-interacting band calculation, is unreliable. We therefore do not use that raw distance alone. We also encode whether a saddle point falls inside a tolerance window of width $w$ that stands in for the interaction scale. The window $w$ is the single knob indexing our nested model classes.
The data and the targets to beat come from machine learning for superconductors. 3DSC - a dataset of superconductors including crystal structures matches SuperCon entries to Materials Project structures through "artificial doping," which links a doped compound to an undoped parent structure. That link is exactly what a rigid-band shift of the filling level needs. S2SNet: A Pretrained Neural Network for Superconductivity Discovery pretrains on about 69k structures. Neither paper uses momentum-space information, and neither tests across chemical families.
The complexity side supplies the methods. Structural Risk Minimization for Learning Nonlinear Dynamics shows that a complexity bound computed over nested model classes can track the U-shaped test-error curve, even though its absolute values are loose. We reuse that nested-class selection, here with descriptor sets as the classes. Improved Outcome Prediction Across Data Sources Through Robust Parameter Tuning shows that internal cross-validation overfits to batch effects specific to each data source, while deliberately simpler choices transfer across studies. Chemical families play the role of studies in our setting. Understanding Contrastive Learning Requires Incorporating Inductive Biases supplies the cautionary analogue: a high-capacity model can drive training loss down by exploiting shortcuts. On their Boolean hypercube task a linear model reaches 99.5% accuracy while an MLP reaches about 74%. We do not claim a new generalization bound. The contribution is an empirical test of where complexity control pays off in materials prediction.
Research Program
For each matched 3DSC entry we compute a DFT band structure on a dense momentum grid and locate critical points of each band, meaning points where the energy gradient vanishes. We keep the saddles, which are the critical points where the band's curvature matrix has eigenvalues of mixed sign. We then shift the filling level rigidly by the electron count implied by the artificial-doping record.
The descriptor vector is about eight numbers. It holds the signed energy offset of the nearest saddle point to the shifted filling level and its indicator for lying within the window $w$. It records the saddle's multiplicity under crystal symmetry. It includes a flatness ratio: the smaller curvature eigenvalue divided by the larger. A near-zero ratio signals a "higher-order" saddle with a stronger, power-law spike in the density of states. It also includes the density of states at the filling level, smoothed over $w$. Finally, it includes a cheap bare charge susceptibility evaluated at the momentum vectors that connect symmetry-related saddles. This is a one-loop quantity, far cheaper than the diagrammatic calculations that need it as input.
The nested hypothesis classes run from small to large. The smallest is saddle descriptors under a sparse linear model. Next come shallow trees on the same descriptors. Next, those descriptors plus composition features. Next, everything plus the real-space DSOAP features of 3DSC. The largest is a pretrained graph network. We select among them three ways and compare the choices: by random-split cross-validation, by leave-one-family-out validation, and by a Rademacher-style complexity penalty in the manner of the Structural Risk Minimization paper.
Baselines
The naive reference point is the formula-only XGBoost model from 3DSC - a dataset of superconductors including crystal structures, which reports a mean squared logarithmic error (MSLE) of 0.776 for transition-temperature regression on 3DSC_MP. We also report a family-mean predictor, which should be hard to beat on random splits and useless out of family.
The strongest regression baseline is that paper's XGBoost with DSOAP structural descriptors, reported at MSLE 0.748 on 3DSC_MP. It is the right competitor because it already shows that structural information regularizes the model. Our saddle features have to add something beyond that real-space information.
For superconductor/non-superconductor classification, the strongest baseline is S2SNet. It reports 91.28% out-of-sample accuracy and an area under the ROC curve (AUC) of 0.9118 on the S2S dataset, against 82.4% for CGCNN and 80.3% for MEGNet. S2SNet is the strongest competitor because its pretraining is the best available way for a high-capacity model to escape memorization.
All baselines will be rerun on identical matched subsets and identical splits, with the same hyperparameter search budget measured in trials. S2SNet keeps its pretraining, and the descriptor models get the same number of tuning trials. We found no reported result on Lacuna for any of these models under leave-one-family-out splits, so those numbers will come only from our reruns. Beating a baseline means a lower held-out-family MSLE, or a higher held-out-family AUC, averaged over families, with paired bootstrap intervals that exclude zero.
Ablations isolate the key ingredient. In the first, we replace all saddle descriptors with the density of states at the filling level alone. In the second, we use the raw saddle offset without the window. In the third, we drop the doping shift. In the fourth, we shuffle the saddle descriptors within each family, which preserves family identity but destroys the physics.
Why not just…?
Why not just add density of states at the filling level to XGBoost? This is the most serious threat, because a spike in states is the mechanism's main effect. The first ablation tests it directly. If the density-of-states-only model matches the full descriptor set out of family, we will report that saddle geometry adds nothing. In that case the useful contribution narrows to the doping shift and the split protocol. We expect the flatness ratio and the saddle-connecting susceptibility to matter for the unconventional families that the physics papers single out. That expectation is a hypothesis, not established.
Why not just use a bigger pretrained graph network? S2SNet's 91.28% accuracy is real. The JARVIS-Leaderboard also finds graph networks ahead of descriptor models on most tasks, though descriptors come within 20% on nearly a third of them. But random splits let family identity leak between train and test. The leave-one-family-out comparison settles the question. If S2SNet still wins there, the low-capacity hypothesis fails.
Why not just run diagrammatic calculations or the functional renormalization group directly? These methods are the right tools for one material. The compensated-metallicity paper notes that the simpler of them, the random phase approximation, overestimates pairing for small Fermi pockets, so the more accurate renormalization-group treatment is needed, at higher cost. Neither is feasible across thousands of SuperCon entries, let alone the 5.3 million entries of LeMat-Bulk. We borrow only their cheapest ingredient, the bare susceptibility.
Why not just apply robust tuning to the existing features? The "Robust1SE" rule from the cross-study tuning paper picks the simplest model whose error is within one standard error of the best. It may recover much of any out-of-family gain on its own. We therefore apply Robust1SE to every baseline as well. The saddle features count as useful only if they win after both sides are tuned this way.
Proposed Experiments
The decisive experiment is a model-by-split comparison on 3DSC_MP regression and on S2S classification. Families are defined by chemistry (cuprates, iron-based, intermetallics, and others), and each family is held out in turn.
The proposal is supported if four things happen together. Random-split cross-validation must choose the graph network or the DSOAP model. Leave-one-family-out validation and the complexity-penalized choice must both choose a saddle-descriptor class. That class must beat the 0.748-MSLE design rerun out of family. And the shuffled-descriptor and density-of-states-only ablations must lose that advantage.
A secondary check sweeps the tolerance window $w$. If interaction pinning matters, held-out error should be minimized at a nonzero $w$, around the scale of the band renormalizations reported for twisted bilayer graphene and NbSe2, rather than at $w \to 0$. Finally, we score kagome compounds and SrPtAs blind, as a qualitative check that the model ranks known saddle-point-driven materials highly. These compounds are too few to support a statistical claim.
References
- Van-Hove tuning of Fermi surface instabilities through compensated metallicity
- Analysis of Charge Order in the Kagome Metal AV3Sb5 (A=K,Rb,Cs)
- Discovery of charge order above room-temperature in the prototypical kagome superconductor La(Ru1−xFex)3Si2
- Chiral<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"><mml:mi>d</mml:mi></mml:math>-wave superconductivity in SrPtAs
- Does filling-dependent band renormalization aid pairing in twisted bilayer graphene?
- Quasiparticle energy bands and Fermi surfaces of monolayer <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"><mml:msub><mml:mi>NbSe</mml:mi><mml:mn>2</mml:mn></mml:msub></mml:math>
- 3DSC - a dataset of superconductors including crystal structures
- S2SNet: A Pretrained Neural Network for Superconductivity Discovery
- Structural Risk Minimization for Learning Nonlinear Dynamics
- Improved Outcome Prediction Across Data Sources Through Robust Parameter Tuning
- Understanding Contrastive Learning Requires Incorporating Inductive Biases
- JARVIS-Leaderboard: a large scale benchmark of materials design methods
- LeMat-Bulk: aggregating, and de-duplicating quantum chemistry materials databases