Search arXivSearch

arXiv · 2503.10471

Siamese Foundation Models for Crystal Structure Prediction

Abstract

Predicting crystal structures from chemical compositions is a fundamental challenge in materials discovery, complicated by complex 3D geometries that distinguish it from fields like protein folding. Here, we present Diffusion-based Crystal Omni (DAO), a pretrain-finetune framework for crystal structure prediction integrating two Siamese foundation models: a structure generator and an energy predictor. The generator is pretrained via a two-stage pipeline on a vast dataset of stable and unstable structures, leveraging the predictor to relax unstable configurations and guide the generative sampling. Across two well-known benchmarks, pretraining significantly enhances performance across multiple backbone architectures. Ablation studies confirm that the synergy between the generator and predictor mutually benefits both components. We further validate DAO on three real-world superconductors ($\text{Cr}_6\text{Os}_2$, $\text{Zr}_{16}\text{Rh}_8\text{O}_4$, and $\text{Zr}_{16}\text{Pd}_8\text{O}_4$) typically inaccessible to conventional computation. For $\text{Cr}_6\text{Os}_2$, DAO achieves a 100\% match rate with experimental references and an atomic-position error of 0.0012 under 20-shot generation, performing over 2000$\times$ faster per iteration than DFT-based structure predictors. These compelling results collectively highlight the potential of our approach for advancing materials science research.

Explore related subjects

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Liming Wu, Wenbing Huang, Rui Jiao, Jianxing Huang, Liwei Liu, Yipeng Zhou, Hao Sun, Yang Liu, Fuchun Sun, Yuxiang Ren, Jirong Wen. 2026-04-14. Siamese Foundation Models for Crystal Structure Prediction. https://arxiv.org/abs/2503.10471

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related papers

Incommensurate structural and magnetic modulations in potassium-rich cryptomelane, K$_x$Mn$_8$O$_{16}$ ($x\approx1.45$)

Cryptomelane is a hollandite-like material consisting of K$^+$ cations in an $α$-MnO$_2$ tunnel-like crystallographic motif. Cryptomelane with stoichiometry K$_x$Mn$_8$O$_{16}$ ($x\approx1.45$) has been synthesized and its magnetic properties investigated using variable-temperature magnetic susceptibility, heat capacity, and neutron powder diffraction. Three distinct transitions at $T_1=184$\,K, $T_2=54.5$\,K, and $T_3=24$\,K are observed. At $T_1$ there is a subtle tetragonal$\rightarrow$monoclinic transition associated with emergence of a set of non-magnetic superstructure peaks indexable to a $\vec{k}_\mathrm{struc}\approx0.74\vec{c^*}$ incommensurate modulation parallel to the $α$-MnO$_2$ tunnels. Our findings are consistent with a relation previously reported in titanate hollandites, that $x\approx2|\vec{k}_\mathrm{struc}|$. Magnetic Bragg peaks emerge below $T_2=54.5$\,K, and their positions indicate an incommensurate modulated magnetic structure. The model consistent with the data is a dual-$\vec{k}_\mathrm{mag}$ structure with a ferromagnetic $|\vec{k}_\mathrm{mag}|=0$ component and an incommensurate $\vec{k}_\mathrm{mag}\approx0.37\vec{c^*}$, with the latter most likely to be helical. The period of oscillation of the incommensurate magnetic component is in line with predictions based on a Heisenberg spin Hamiltonian [Mandal \textit{et al}. Phys. Rev. B 90, 104420 (2014)]. Below $T_3=24$\,K, there is a magnetic transition, which gives rise to a different set of magnetic Bragg peaks indicative of a highly complex magnetic structure.

cond-mat.mtrl-sci

An anisotropic functional for two-dimensional material systems

Density function theory is the workhorse of modern electronic structure theory. However, its accuracy in practical calculations is limited by the choice of the exchange-correlation potential. In this respect, two-dimensional materials pose a special challenge, as all these materials and their heterostructures have a crucial similarity. The underlying atomic structures are strongly spatially inhomogeneous, implying that current exchange-correlation functionals, that in almost all cases are isotropic, are ill-prepared for an accurate description. We present an anisotropic screened-exchange potential, that remedies this problem and reproduces the band-gap of 2D materials as well as the piecewise linearity of the total energy with fractional occupation number.

cond-mat.mtrl-sci

Thermally-driven reorientation of the Néel vector in altermagnetic MnTe

Altermagnets are novel magnetic systems that possess a spin-polarized electronic band structure without a net magnetic moment, making them promising for device applications. Hexagonal MnTe, a prototypical altermagnet, arguably exhibits the most properties consistent with theoretical predictions, including an anomalous Hall effect despite a vanishing net magnetization, and altermagnetinduced electronic band splitting. However, fundamental questions remain, including why some effects only appear significantly below the magnetic ordering temperature. Here, we resolve this discrepancy by revealing a reorientation of the Néel vector in single-crystalline MnTe. The Néel vector points 30° from the a-axis at low $T$, before aligning directly with the a-axis around $T\simeq 260$ K. We attribute this to single-ion anisotropy, which depends on temperature-dependent lattice parameters. We obtained these results using muon-spin spectroscopy, magnetization measurements, and X-ray diffraction; we show that the findings are consistent with neutron diffraction. Manipulating this effect, for example through strain, could unlock sensitive electronic detection schemes for external stimuli, paving the way for functional altermagnetic devices.

cond-mat.mtrl-sci