Searcharxiv⌕ Search

arXiv subjects

Yizhi Zhu

Publications and source records attributed to Yizhi Zhu.

6 recordsLinked to original sources

Worse Forces Can Lead to Better Structures: Trajectory-Level Fine-Tuning of Machine Learning Interatomic Potentials

Machine learning interatomic potentials (MLIPs) are commonly trained and evaluated by their ability to reproduce first-principles energies and forces. Practical value, however, is often determined by downstream processes such as structural relaxation. Much of MLIP literature is motivated by the expectation that improving force accuracy will produce correspondingly better downstream behavior. Although this expectation is natural, it may not hold in practice due to model inaccuracies. Here, we demonstrate that force accuracy and relaxation accuracy can be driven in opposite directions. Starting from pretrained MLIPs, we differentiate through relaxation trajectories and optimize the the final structures. This intervention produces increases in force-prediction errors while substantially reducing structural errors. Importantly, this represents a trade-off between two aspects of physical fidelity: the learned forces become less faithful to first-principles forces, while the resulting relaxed structures become more faithful to first-principles predictions. Rather than observing a discrepancy between local quality and downstream performance, we demonstrate that this discrepancy can be induced through optimization. Our findings suggest that MLIPs need not be optimized as approximations of first-principles forces, but can instead learn task-specific vector fields that guide relaxation trajectories toward more accurate final structures. By propagating supervision through complete relaxation trajectories, our framework captures dependencies between successive predictions that are absent when structure--force pairs are treated independently. More broadly, these results demonstrate that conventional force metrics can provide an incomplete assessment of downstream performance and highlights the importance of evaluating scientific machine-learning models within computational workflows.

cond-mat.mtrl-sci↗

DefectSkills: Guiding AI agents through first-principles calculations for point defects in semiconductors and insulators

First-principles defect calculations can now be performed autonomously by AI agents, enabled by established computational methods and open-source Python toolkits. However, reliable and repeatable defect calculations require clear scientific and technical guidance that helps agents resolve ambiguities and avoid errors in calculation setup, execution, postprocessing, and analysis. Here, we present DefectSkills, a set of structured prompts that provides human-in-the-loop guidance to AI agents performing first-principles calculations of point defects in semiconductors and insulators using the Vienna Ab initio Simulation Package (VASP). The prompts encode human expertise and guide agents through unit-cell relaxation, electronic-structure and dielectric calculations, phase-stability and chemical-potential analysis, charged-defect supercell generation and calculations, finite-size corrections, formation-energy calculations and thermodynamic analysis, and data backup. By packaging this expertise as reusable agent skills, DefectSkills supports consistent workflow execution across agentic AI systems and reproducible defect calculations.

cond-mat.mtrl-sci↗

ADAPT: Lightweight, Long-Range Machine Learning Force Fields Without Graphs

Point defects play a central role in driving the properties of materials. First-principles methods are widely used to compute defect energetics and structures, including at scale for high-throughput defect databases. However, these methods are computationally expensive, making machine-learning force fields (MLFFs) an attractive alternative for accelerating structural relaxations. Most existing MLFFs are based on graph neural networks (GNNs), which can suffer from oversmoothing, oversquashing, and poor representation of long-range interactions. Both of these issues are especially of concern when modeling point defects. To address these challenges, we introduce the \textit{Accelerated Deep Atomic Potential Transformer} (ADAPT), an MLFF that replaces graph representations with a direct coordinates-in-space formulation and explicitly considers all pairwise atomic interactions. Atoms are treated as tokens, with a Transformer encoder modeling their interactions. Applied to a dataset of silicon point defects, ADAPT achieves a roughly 22% reduction in force and a roughly 40 percent reduction in energy prediction error relative to a state-of-the-art GNN-based model, while requiring only a fraction of the computational cost.

cs.LG↗

Identifying high performance spectrally-stable quantum defects in diamond

Point defects in semiconductors are becoming central to quantum technologies. They can be used as spin qubits interfacing with photons, which are fundamental for building quantum networks. Currently, the most prominent quantum defect in diamond is the nitrogen-vacancy (NV) center. However, it suffers from spectral diffusion that negatively impacts optical coherence and is due to the coupling of the emission energy with uncontrolled electric fields. The group IV vacancy complexes on the other hand have shown to be significantly more spectrally-stable as they are centrosymmetric and thus immune to the linear Stark shift. They however suffer from several issues ranging from low operation temperature to low optical efficiency due to dark states and difficulty in stabilizing the right defect charge state. Here we search for alternative to the group IV vacancy complex in diamond by systematically evaluating all possible vacancy complex using high-throughput first-principles computational screening. We identify the defects that combine centrosymmetry, emission in the visible range, as well as favorable and achievable electronic structure promoting higher operation temperature and defect levels well within the band gap. We find Zn$V^{-2}$ to be especially appealing.

cond-mat.mtrl-sci↗

Quantum spin probe of single charge dynamics

Electronic defects in semiconductors form the basis for many emerging quantum technologies. Understanding defect spin and charge dynamics in solid state platforms is crucial to developing these building blocks, but many defect centers are difficult to access at the single-particle level due to the lack of sensitive readout techniques. A method for probing optically inactive spin defects would reveal semiconductor physics at the atomic scale and advance the study of new quantum systems. We exploit the intrinsic correlation between the charge and spin states of defect centers to measure defect charge populations and dynamics through the steady-state spin population, read-out at the single-defect level with a nearby optically active qubit. We directly measure ionization and charge relaxation of single dark defects in diamond, effects we do not have access to with traditional coherence-based quantum sensing. These spin resonance-based methods generalize to other solid state defect systems in relevant materials.

quant-ph↗

Five-second coherence of a single spin with single-shot readout in silicon carbide

An outstanding hurdle for defect spin qubits in silicon carbide (SiC) is single-shot readout - a deterministic measurement of the quantum state. Here, we demonstrate single-shot readout of single defects in SiC via spin-to-charge conversion, whereby the defect's spin state is mapped onto a long-lived charge state. With this technique, we achieve over 80% readout fidelity without pre- or post-selection, resulting in a high signal-to-noise ratio (SNR) that enables us to measure long spin coherence times. Combined with pulsed dynamical decoupling sequences in an isotopically purified host material, we report single spin T2 > 5s, over two orders of magnitude greater than previously reported in this system. The mapping of these coherent spin states onto single charges unlocks both single-shot readout for scalable quantum nodes and opportunities for electrical readout via integration with semiconductor devices.

quant-ph↗