npj Computational Materials最新文献

In this work, a multi-stage material design framework combining machine learning techniques with density functional theory is established to reveal the mechanism of phase stabilization in HfO₂ based ferroelectric materials. The ferroelectric phase fractions based on a more stringent relationship of phase energy differences is proposed as an evaluation criterion for the ferroelectric performance of hafnium-based materials. Based on the Boltzmann distribution theory, the abstract phase energy difference is converted into an intuitive phase fraction distribution mapping. A large-scale prediction of unknown dopants is conducted within the material design framework, and gallium (Ga) is identified as a new dopant for HfO₂. Both experiments and density functional theory calculations demonstrate that Ga is an excellent dopant for ferroelectric hafnium oxide, especially, the experimentally determined variation trends of ferroelectric phase fraction and polarization properties with Ga doping concentration are in good agreement with the predictions given by machine learning. This work provides a new perspective from machine learning to deepen the understanding of the ferroelectric properties of HfO₂ materials, offering fresh insights into the design and performance prediction of HfO₂ ferroelectric thin films.

A simulation can stand its ground against an experiment only if its prediction uncertainty is known. The unknown accuracy of interatomic potentials (IPs) is a major source of prediction uncertainty, severely limiting the use of large-scale classical atomistic simulations in a wide range of scientific and engineering applications. Here we explore covariance between predictions of metal plasticity, from 178 large-scale (~10⁸ atoms) molecular dynamics (MD) simulations, and a variety of indicator properties computed at small-scales (≤10² atoms). All simulations use the same 178 IPs. In a manner similar to statistical studies in public health, we analyze correlations of strength with indicators, identify the best predictor properties, and build a cross-scale “strength-on-predictors” regression model. This model is then used to estimate regression error over the statistical pool of IPs. Small-scale predictors found to be highly covariant with strength are computed using expensive quantum-accurate calculations and used to predict flow strength, within the statistical error bounds established in our study.

A pathway to overcome limited data availability in materials science is to use the framework of transfer learning, where a pre-trained (PT) machine learning model (on a larger dataset) can be fine-tuned (FT) on a target (smaller) dataset. We systematically explore the effectiveness of various PT/FT strategies to learn and predict material properties and create generalizable models by PT on multiple properties (MPT) simultaneously. Specifically, we leverage graph neural networks (GNNs) to PT/FT on seven diverse curated materials datasets, with sizes ranging from 941 to 132,752. Besides identifying optimal PT/FT strategies and hyperparameters, we find our pair-wise PT-FT models to consistently outperform models trained from scratch on target datasets. Importantly, our MPT models outperform pair-wise models on several datasets and, more significantly, on a 2D material band gap dataset that is completely out-of-domain. Finally, we expect our PT/FT and MPT frameworks to accelerate materials design and discovery for various applications.

Stable or metastable crystal structures of assembled atoms can be predicted by finding the global or local minima of the energy surface within a broad space of atomic configurations. Generally, this requires repeated first-principles energy calculations, which is often impractical for large crystalline systems. Here, we present significant progress toward solving the crystal structure prediction problem: we performed noniterative, single-shot screening using a large library of virtually created crystal structures with a machine-learning energy predictor. This shotgun method (ShotgunCSP) has two key technical components: transfer learning for accurate energy prediction of pre-relaxed crystalline states, and two generative models based on element substitution and symmetry-restricted structure generation to produce promising and diverse crystal structures. First-principles calculations were performed only to generate the training samples and to refine a few selected pre-relaxed crystal structures. The ShotunCSP method is less computationally intensive than conventional methods and exhibits exceptional prediction accuracy, reaching 93.3% in benchmark tests with 90 different crystal structures.

Koopmans spectral functionals are a powerful extension of Kohn-Sham density-functional theory (DFT) that enables the prediction of spectral properties with state-of-the-art accuracy. The success of these functionals relies on capturing the effects of electronic screening through scalar, orbital-dependent parameters. These parameters have to be computed for every calculation, making Koopmans spectral functionals more expensive than their DFT counterparts. In this work, we present a machine-learning model that—with minimal training—can predict these screening parameters directly from orbital densities calculated at the DFT level. We show in two prototypical use cases that using the screening parameters predicted by this model, instead of those calculated from linear response, leads to orbital energies that differ by less than 20 meV on average. Since this approach dramatically reduces run times with minimal loss of accuracy, it will enable the application of Koopmans spectral functionals to classes of problems that previously would have been prohibitively expensive, such as the prediction of temperature-dependent spectral properties. More broadly, this work demonstrates that measuring violations of piecewise linearity (i.e., curvature in total energies with respect to occupancies) can be done efficiently by combining frozen-orbital approximations and machine learning.

The impact of magnetism on chemical ordering in face-centered cubic CrCoNi medium entropy alloy is studied by a combination of ab initio simulations, machine learning potentials, and Monte Carlo simulations. Large magnetic energies are revealed for some mixed L1₂/L1₀ type ordered configurations, which are rooted in strong nearest-neighbor magnetic exchange interactions and chemical bonding among the constituent elements. There is a delicate interplay between magnetism and stability of MoPt₂ and L1₂/L1₀ type of order, which may explain opposing experimental and theoretical findings.

We developed a general framework for hybrid quantum-classical computing of molecular and periodic embedding approaches based on an orbital space separation of the fragment and environment degrees of freedom. We demonstrate its potential by presenting a specific implementation of periodic range-separated DFT coupled to a quantum circuit ansatz, whereby the variational quantum eigensolver and the quantum equation-of-motion algorithm are used to obtain the low-lying spectrum of the embedded fragment Hamiltonian. The application of this scheme to study localized electronic states in materials is showcased through the accurate prediction of the optical properties of the neutral oxygen vacancy in magnesium oxide (MgO). Despite some discrepancies in the position of the main absorption band, the method demonstrates competitive performance compared to state-of-the-art ab initio approaches, particularly evidenced by the excellent agreement with the experimental photoluminescence emission peak.

Achieving room-temperature superconductivity under ambient conditions is one of the most important goals in solid-state physics and material science. Recent discoveries of high-T_c superconductivity in binary hydrides H₃S and LaH₁₀ at high pressures have focused the search for room-temperature superconductors on dense hydrides with conventional phonon-mediated pairing mechanisms. In this study, we predict a novel family of superconducting ternary hydrides under moderate compression, XC₂H₈ (X = Ga, In, Tl, Sn, Pb, Sb, Bi, Te, Po). Unlike H₃S and LaH₁₀, these new materials are stable at just around 20 GPa. Among the analyzed compounds, SbC₂H₈ exhibits the highest critical temperature of 73 K at a pressure of 100 GPa, which is attributed to its energetically favorable high-symmetry crystal structure ((Fm{bar{3}}m)), high density of states at the Fermi level (1.27 states/eV) and strong electron–phonon coupling constant (1.02). We expect that our findings provide crucial insights into achieving high-temperature superconductivity at moderate pressures and accelerate the progress of experimental research.

A major obstacle to the realization of novel inorganic materials with desirable properties is efficient materials discovery over both the materials property and synthesis spaces. In this work, we propose and compare two novel reinforcement learning (RL) approaches to inverse inorganic oxide materials design to target promising compounds using specified property and synthesis objectives. Our models successfully learn chemical guidelines such as negative formation energy, charge neutrality, and electronegativity balance while maintaining high chemical diversity and uniqueness. We demonstrate multi-objective RL algorithms that can generate novel compounds with both desirable materials properties (band gap, formation energy, bulk modulus, shear modulus) and synthesis objectives (low sintering and calcination temperatures). We apply template-based crystal structure prediction to suggest feasible crystal structure matches for target inorganic compositions identified by our machine learning (ML) algorithms to highlight the plausibility of the identified target compositions. We analyze the benefits and drawbacks of the ML approaches tested in this work in the context of accelerated inorganic materials design. This work isolates and evaluates the effects of different RL methodologies to suggest promising, valid compounds of interest by exploring the chemical design space for materials discovery.

Solid polymer electrolytes offer promising advancements for next-generation batteries, boasting superior safety, enhanced specific energy, and extended lifespans over liquid electrolytes. However, low ionic conductivity and the vast polymer space hinder commercialization. This study leverages generative AI for de novo polymer electrolyte design, comparing GPT-based and diffusion-based models with extensive hyperparameter tuning. We evaluate these models using various metrics and full-atom molecular dynamics simulations. Among 46 candidates tested, 17 exhibit superior ionic conductivity, surpassing existing polymers in our database, with some doubling the conductivity values. Additionally, by adopting pretraining and fine-tuning methodologies, we significantly enhance our generative models, achieving quicker convergence, better performance with limited data, and greater diversity. Our method efficiently generates a large number of novel, diverse, and valid polymers, with a high likelihood of synthesizability, enabling the identification of promising candidates with markedly improved efficiency and effectiveness for practical applications.