We introduce CrystalFormer, a transformer-based autoregressive model specifically designed for space group-controlled generation of crystalline materials. By explicitly incorporating space group symmetry, CrystalFormer greatly reduces the effective complexity of crystal space, which is essential for data-and compute-efficient generative modeling of crystalline materials. Leveraging the prominent discrete and sequential nature of the Wyckoff positions, CrystalFormer learns to generate crystals by directly predicting the species and coordinates of symmetry-inequivalent atoms in the unit cell. We demonstrate the advantages of CrystalFormer in standard tasks such as symmetric structure initialization and element substitution over widely used conventional approaches. Furthermore, we showcase its plug-and-play application to property-guided materials design, highlighting its flexibility. Our analysis reveals that CrystalFormer ingests sensible solid-state chemistry knowledge and heuristics by compressing the material dataset, thus enabling systematic exploration of crystalline materials space. The simplicity, generality, and adaptability of CrystalFormer position it as a promising architecture to be the foundational model of the entire crystalline materials space, heralding a new era in materials discovery and design.

Recent advances in deep learning generative models (GMs) have created high capabilities in accessing and assessing complex high-dimensional data, allowing superior efficiency in navigating vast material configuration space in search of viable structures. Coupling such capabilities with physically significant data to construct trained models for materials discovery is crucial to moving this emerging field forward. Here, we present a universal GM for crystal structure prediction (CSP) via a conditional crystal diffusion variational autoencoder (Cond-CDVAE) approach, which is tailored to allow user-defined material and physical parameters such as composition and pressure. This model is trained on an expansive dataset containing over 670,000 local minimum structures, including a rich spectrum of high-pressure structures, along with ambient-pressure structures in Materials Project database. We demonstrate that the Cond-CDVAE model can generate physically plausible structures with high fidelity under diverse pressure conditions without necessitating local optimization, accurately predicting 59.3% of the 3547 unseen ambient-pressure experimental structures within 800 structure samplings, with the accuracy rate climbing to 83.2% for structures comprising fewer than 20 atoms per unit cell. These results meet or exceed those achieved via conventional CSP methods based on global optimization. The present findings showcase substantial potential of GMs in the realm of CSP.

The rapid advancements in artificial intelligence (AI) are catalyzing transformative changes in atomic modeling, simulation, and design. AI-driven potential energy models have demonstrated the capability to conduct large-scale, long-duration simulations with the accuracy of ab initio electronic structure methods. However, the model generation process remains a bottleneck for large-scale applications. We propose a shift towards a model-centric ecosystem, wherein a large atomic model (LAM), pre-trained across multiple disciplines, can be efficiently fine-tuned and distilled for various downstream tasks, thereby establishing a new framework for molecular modeling. In this study, we introduce the DPA-2 architecture as a prototype for LAMs. Pre-trained on a diverse array of chemical and materials systems using a multi-task approach, DPA-2 demonstrates superior generalization capabilities across multiple downstream tasks compared to the traditional single-task pre-training and fine-tuning methodologies. Our approach sets the stage for the development and broad application of LAMs in molecular and materials simulation research.

Crystal structure prediction (CSP) and machine learning potential (MLP) are two fundamental methods for modern computational material discovery. While the former aims at efficient sampling of the potential energy surface (PES) for discovering new materials, the latter focuses on reproducing the PES to accelerate various atomic simulation tasks. In this work, we combine the two methods within a concurrent learning framework in an effort to generate efficient MLP models for accelerating CSP. The proposed scheme explores the PES through the swarm-intelligence calypso method, labels the most representative structures with quantum mechanical calculations, and learns the PES through a deep potential (DP) model. The process proceeds in an iterative, computationally efficient, and automated manner, leading to the collection of a most compact reference training set from which the resulting DP model is proven particularly suitable for accelerating calypso structure prediction. The scheme has been systematically benchmarked on binary magnesium-aluminium (Mg-Al) alloys and ternary lithium-lanthanum-hydrogen (Li-La-H) superhydrides, demonstrating its efficiency and reliability in DP model construction and calypso structure prediction. The proposed scheme represents a promising routine to perform the structure prediction of large or multicomponent systems.

The rapid development of artificial intelligence (AI) is driving significant changes in the field of atomic modeling, simulation, and design. AI-based potential energy models have been successfully used to perform large-scale and long-time simulations with the accuracy of ab initio electronic structure methods. However, the model generation process still hinders applications at scale. We envision that the next stage would be a model-centric ecosystem, in which a large atomic model (LAM), pre-trained with as many atomic datasets as possible and can be efficiently fine-tuned and distilled to downstream tasks, would serve the new infrastructure of the field of molecular modeling. We show that DPA-2 can accurately represent a diverse range of chemical systems and materials, enabling high-quality simulations and predictions with significantly reduced efforts compared to traditional methods. Our approach paves the way for a universal large atomic model that can be widely applied in molecular and material simulation research, opening new opportunities for scientific discoveries and industrial applications.

Crystal structure prediction is a powerful theoretical simulation tool, which can determine the crystal structure of materials with the given information of chemical composition. However, its application is severely limited due to the highly computational cost. In recent years, the state-of-art machine learning methods reveal a promising prospect in accelerating the conventional scientific computing, thus introducing the methods into the crystal structure prediction. This review briefly introduced recent progress on the application of machine learning for the crystal structure prediction. Two aspects were discussed, i.e., accelerating the energy evaluation and enhancing the potential energy surface sampling. In addition, some insights into the future development in this aspect were also suggested.

Alkaline earth metal peroxides are typical examples of ionic compounds containing polyanions. We herein report a stable $\mathrm{Ba}{\mathrm{O}}_{2}$ phase at high pressure up to 130 GPa found via a first-principles computational structure search and high-pressure experimental investigations. The identified monoclinic structure (space group $C2/m$) can be derived by sublattice distortions of Ba atoms and peroxide groups associated with the phonon mode softening of the lower-pressure $Cmmm$ structure. Contrary to the previous expectation of polymerization of the peroxide group at elevated pressure, this phase retains the peroxide group and, interestingly, exhibits an insulating behavior demonstrating an increase of the band gap under compression. Our synchrotron x-ray diffraction (XRD) measurements could not distinguish between $Cmmm$ and $C2/m\mathrm{Ba}{\mathrm{O}}_{2}$ definitively because the difference in XRD patterns is very subtle. However, our data do not show any sign of polymerization transition up to 120 GPa. Raman spectra of the O-O peroxide vibration show a small anomaly in frequency at 110 GPa, which is qualitatively like that predicted theoretically due to the $Cmmm$ to $C2/m$ phase transition, thus supporting the predicted transformation.

Simulating reconstructive phase transition requires an accurate description of potential energy surface (PES). Density-functional-theory (DFT) based molecular dynamics can achieve the desired accuracy but it is computationally unfeasible for large systems and/or long simulation times. Here we introduce an approach that combines the metadynamics simulation and machine learning representation of PES at the accuracy close to the DFT calculations, but with the computational cost several orders of magnitude less, and scaling with system size approximately linear. The high accuracy of the method is demonstrated in the simulation of pressure-induced $B4\text{\ensuremath{-}}B1$ phase transition in gallium nitride (GaN). The large-scale simulation using a 4096-atom simulation box reveals the phase transition with excellent detail, revealing different simulated transition paths under particular stress conditions. With well-trained machine learning potentials, this method can be easily applied to all types of systems for accurate scalable simulations of solid-solid reconstructive phase transition.