Abstract

Crystal Structure Prediction (CSP), which aims to generate stable crystalstructures from compositions, represents a critical pathway for discoveringnovel materials. While structure prediction tasks in other domains, such asproteins, have seen remarkable progress, CSP remains a relatively underexploredarea due to the more complex geometries inherent in crystal structures. In thispaper, we propose Siamese foundation models specifically designed to addressCSP. Our pretrain-finetune framework, named DAO, comprises two complementaryfoundation models: DAO-G for structure generation and DAO-P for energyprediction. Experiments on CSP benchmarks (MP-20 and MPTS-52) demonstrate thatour DAO-G significantly surpasses state-of-the-art (SOTA) methods across allmetrics. Extensive ablation studies further confirm that DAO-G excels ingenerating diverse polymorphic structures, and the dataset relaxation andenergy guidance provided by DAO-P are essential for enhancing DAO-G'sperformance. When applied to three real-world superconductors($\text{CsV}_3\text{Sb}_5$, $ \text{Zr}_{16}\text{Rh}_8\text{O}_4$ and$\text{Zr}_{16}\text{Pd}_8\text{O}_4$) that are known to be challenging toanalyze, our foundation models achieve accurate critical temperaturepredictions and structure generations. For instance, on$\text{CsV}_3\text{Sb}_5$, DAO-G generates a structure close to theexperimental one with an RMSE of 0.0085; DAO-P predicts the $T_c$ value withhigh accuracy (2.26 K vs. the ground-truth value of 2.30 K). In contrast,conventional DFT calculators like Quantum Espresso only successfully derive thestructure of the first superconductor within an acceptable time, while the RMSEis nearly 8 times larger, and the computation speed is more than 1000 timesslower. These compelling results collectively highlight the potential of ourapproach for advancing materials science research and development.