Stephan Lany

Oxygen vacancy formation energies govern the performance of metal oxides across energy conversion, catalysis, and electronics, yet predicting them accurately without density functional theory calculations for novel chemical systems remains challenging. Here, we show that frozen 256-dimensional embeddings extracted from a pretrained MACE universal machine-learning interatomic potential encode information sufficient to predict vacancy formation energies obtained from density functional theory without requiring supercell construction, vacancy creation, or geometry optimization. A lightweight multilayer perceptron trained on these embeddings (MACE-dGNN) achieves an element-wise cross-validation mean absolute error of 0.38 eV, halving the error of a baseline graph neural network trained from scratch (0.72 eV) and offering a modest advantage over direct MACE relaxation calculations (0.51 eV). Because the embedding approach decouples representation from the prediction task and generalizes well with limited data, it will (1) extend naturally to fine-tuning and prediction of properties (e.g., charged defects) that are inaccessible to generic energy-and force-output interatomic potentials and (2) be trainable on necessarily small datasets, potentially enabling the use of more expensive but accurate tools for generating training data (e.g., hybrid functional calculations) for which training or fine-tuning of interatomic potentials will be difficult. We demonstrate practical impact by revisiting a prior screening of thermochemical water-splitting materials, where improved generalization alters ∼45 % of candidate classifications.

Incorporation of AlyGa1–yN semiconductors into high-power electronics offers efficiency improvements in power transmission, generation, and use, if approaches to eliminate the defects arising from film-lattice mismatch can be established. Here, we report the optical floating zone crystal growth of Ta1–xHfxC0.5 (x = 0.2), a new metallic substrate material family lattice matched to the ultrawide-band gap, Al-rich side (y = 0.91) of the AlyGa1–yN solid solution. Laue diffraction demonstrates large single-crystal domains in the as-grown boule. Single-crystal X-ray diffraction at T = 213 K in conjunction with first-principles calculations shows that the material adopts a layered crystal structure with AA-type stacking of (Ta/Hf)-C-(Ta/Hf) trilayers described in the trigonal space group P3̅m1 (#164), with a = 3. 1168(4) Å, c = 4.9644(4) Å, and β = 120.0°. X-ray photoelectron spectroscopy (XPS) measurements show the Hf:Ta ratio to be close to the nominal value of 0.8:0.2 in the grown crystal. Density Functional Theory calculations reveal that this structure is stabilized by the low energy of carbon-vacancy formation of a hypothetical (Ta/Hf)1C1 anti-NiAs structure type, and imply flexibility in interface structure with an overlayer nitride film. A surface preparation/polishing procedure is developed that reduces root-mean-square (RMS) surface roughness from as-cut 130 to 7 nm as measured by atomic force microscopy. Scanning electron microscopy shows the presence of a native surface oxide, removed by polishing, along with carbon-rich pits. Time-domain thermoreflectance measurements show a room temperature thermal conductivity of κ = 18.1(4) W m–1 K–1. These results provide key first steps for utilizing metallic, lattice-matched substrates for the growth of Al-rich AlyGa1–yN semiconductors.

Hexagonal perovskite oxides with 12R stacking host well-separated face-sharing octahedral metal trimers with short metal–metal distances, leading in some cases to large degrees of magnetic frustration and cluster magnetism. Introducing magnetic ions in proximity to these trimers can influence the degree of frustration and cluster magnetism, but the extent to which the magnetism can be tuned by varying neighboring metal cations remains an open question. In this work, we test the impacts of using Nd in proximity to Mn trimers in the hexagonal perovskite 12R-Ba4NdMn3O12−γ. Given this stoichiometry, Nd should assume the 4+ oxidation state with a spin state of S = 1, which would be the first realization of Nd4+ in an oxide environment. Through detailed bulk magnetic, X-ray absorption spectroscopic, and powder neutron diffraction (PND) measurements, we find that Nd4+ is realized in this material, but there is also partial reduction to Nd3+, which is charge-balanced by O vacancies. Magnetometry measurements indicate an antiferromagnetic ordering temperature TN ≈ 16 K, and PND measurements reveal a surprising collinear antiferromagnetic structure with magnetic space group Pc2/m (no. 10.49 in BNS notation), which has not previously been seen in this class of materials. Our results represent a comprehensive analysis of the structural, electronic, and magnetic properties of 12R-Ba4NdMn3O12−γ, showing the first observation of partial Nd4+ in an oxide, and demonstrating that this structural class can host a broad range of magnetic structures which are not easily predicted based on compositional trends.

Thermochemical hydrogen production (TCH) is a promising approach to convert water into hydrogen by utilizing thermal energy from sources such as concentrated solar energy, as opposed to using electricity in electrochemical methods. This process consists of two main steps: first, a material undergoes thermal reduction at elevated temperatures (~ 1300 - 1500 °C) under low oxygen partial pressures (pO₂) of ~10^-3 -10 ^-5 atm, which results in the formation of oxygen vacancies. Next, the temperature of the system is reduced to ~800 - 1000 °C in the presence of steam, enabling water splitting through material oxidation. The commercial feasibility of TCH is dependent on identifying a material with an optimal reduction enthalpy that supports both the thermal reduction of redox-active cations and their re-oxidation, while also possessing high cycle stability. Perovskite metal oxides have received significant interest as potential TCH materials due to their ability to accommodate high concentrations of oxygen vacancies and their highly tunable thermodynamic properties. However, these materials often exhibit low hydrogen production due to re-oxidation and poor cycle stability challenges.

Thermochemical hydrogen (TCH) via water-splitting provides a promising technology pathway for hydrogen production since it, in contrast to electrolysis, does not depend primarily on redirecting electricity from the grid for fuel production. Especially due to recent commercialization efforts, 2-step thermal redox cycles in non-stoichiometric metal oxides are of particularly high interest for this pathway; however, state-of-the-art CeO₂ has several practical limitations, which has motivated continued materials discovery efforts in this field. Our first contribution demonstrates how machine learning models can accelerate the high-throughput screening of metal oxides’ oxygen defect thermodynamics to identify promising novel TCH candidates. Upon their experimental validation, some materials exhibit TCH capabilities comparable to CeO₂ under certain reactor operating conditions. Shifting gears, we then discuss how liquid metal-mediated thermochemical redox can serve as a promising alternative approach due its drastically reduced operating temperatures and promising technoeconomic outlook. Here, machine learned interatomic potentials are instead utilized, accelerating the molecular dynamics simulations needed to understand the phenomena and design rules underpinning their excellent water-splitting capabilities.

Oxygen vacancy formation energies govern the performance of metal oxides across energy conversion, catalysis, and electronics, yet predicting them accurately without density functional theory calculations for novel chemical systems remains challenging. Here, we show that frozen 256-dimensional embeddings extracted from a pretrained MACE universal machine-learning interatomic potential encode information sufficient to predict vacancy formation energies obtained from density functional theory without requiring supercell construction, vacancy creation, or geometry optimization. A lightweight multilayer perceptron trained on these embeddings (MACE-dGNN) achieves an element-wise cross-validation mean absolute error of 0.38 eV, halving the error of a baseline graph neural network trained from scratch (0.72 eV) and offering a modest advantage over direct MACE relaxation calculations (0.51 eV). Because the embedding approach decouples representation from the prediction task and generalizes well with limited data, it will (1) extend naturally to fine-tuning and prediction of properties (e.g., charged defects) that are inaccessible to generic energy-and force-output interatomic potentials and (2) be trainable on necessarily small datasets, potentially enabling the use of more expensive but accurate tools for generating training data (e.g., hybrid functional calculations) for which training or fine-tuning of interatomic potentials will be difficult. We demonstrate practical impact by revisiting a prior screening of thermochemical water-splitting materials, where improved generalization alters ∼45 % of candidate classifications.

Antiperovskite nitrides with the general formula M3AN have attracted significant attention due to their tunable electronic and magnetic properties. Among them are many cobalt-based compounds predicted to exhibit high thermodynamic stability and intriguing magnetic behavior. Here, we report the synthesis and magnetic characterization of epitaxial Co3ZnN thin films grown by radio frequency sputtering on SrTiO3 (STO) and MgO substrates. X-ray diffraction confirms phase-pure (00l)-oriented films with cube-on-cube epitaxy on STO, with a c-lattice parameter of 3.752 Å. Magnetic measurements reveal clear hysteresis at 2 K with a coercive field of ∼0.11 T and a small net moment of 0.108 μB/f.u., suggesting either a canted antiferromagnetic (AFM) or ferrimagnetic (FiM) configuration. Temperature-dependent magnetization measurements show a transition near 25 K, with strong AFM interactions with Curie–Weiss temperature (Θ) = −80.13 K. Complementary density functional theory and Monte Carlo simulations indicate a ferromagnetic (FM) ground state, with the FM–AFM energy difference decreasing systematically with increasing supercell size, consistent with competition between FM and AFM/FiM interactions. These results highlight Co3ZnN as a magnetically complex antiperovskite nitride with competing exchange interactions.

Advances in ultra-wide bandgap materials, such as high Al-content AlxGa1−xN (AlGaN), are essential for next generation power electronics, but the requirement for lattice matched substrates is currently a significant obstacle. Recently, conductive TaC has emerged as a promising virtual substrate for AlGaN heteroepitaxy, with wurtzite (0001) AlxGa1−xN lattice-matched to rocksalt (111) TaC at x ≈ 0.5. Thus, understanding and controlling the electronic properties of the TaC/AlGaN interface is key for developing technological applications based on TaC/AlGaN devices. Using density functional theory and electronic structure calculations, we here investigate TaC/Al0.5Ga0.5N interfaces, where we include explicit alloy models in the slab calculations. We predict the Schottky barrier height and the electric field discontinuity resulting from interface charges. Considering all possible combinations of (Ta, C) substrate termination, (Al/Ga, N) nucleation, and (Al/Ga, N) polarity, we construct a chemical potential phase diagram to identify the stable interfaces that can be accessed through variation of the synthesis conditions. The predicted interface electronic properties are implemented in device performance simulations to demonstrate a practical design for a strain-free, high-efficiency TaC/AlGaN Schottky diode with a low barrier height and without interface charges, underscoring the potential of TaC as a substrate for ultra-wide bandgap devices.