Does the Yield Point of a Glass Depend on Getting the Electrons Right?

When a glass is sheared back and forth, it has two very different long-run behaviors. At small amplitudes, every atom returns to exactly where it started after each cycle: the material is reversible and "remembers" how it was driven. Above a critical amplitude, atoms stop returning. They wander away diffusively, and the glass yields and flows. This switch from localized to diffusive motion is one of the cleanest dynamical transitions in disordered matter. Nearly everything known about it, however, comes from toy glasses: atoms modeled as spheres interacting through fixed pairwise forces. Real metallic, covalent, and oxide glasses bond through electrons. Their bonds bend, rehybridize, and break in ways a pairwise law cannot describe. Density functional theory (DFT), the standard quantum-mechanical method for computing the energy of a set of atoms from their electron density, describes this bonding well. It is far too expensive to run for the hundreds of shear cycles the transition requires.

This note proposes to measure whether the critical amplitude, and the preparation-dependent "threshold energy" that separates gradual from brittle yielding, actually change when the forces come from DFT rather than from a classical formula. The key point is methodological. The quantity that sets yield is the curvature of the energy landscape just before a local minimum vanishes and atoms rearrange. Those near-unstable configurations are exactly the ones that machine-learned DFT potentials, trained on relaxed or thermal structures, have never been checked on. We therefore build DFT-trained potentials whose training data is aimed at these instability points. We validate them against direct DFT shear on small cells, and only then run large-scale cyclic shear on a metallic glass and a covalent glass.

Core Claim

The claim has two parts, and both can fail. First, a potential can reproduce DFT energies of relaxed amorphous structures to tens of meV per atom and still misplace the strain at which plastic events occur. The reason is that yield depends on the lowest curvature of the energy surface near an instability, not on average energies. Second, once the potential is fixed at these instability points, the critical amplitude $\gamma_c$ and the threshold energy of a covalent glass will differ measurably from those given by the best classical potential for the same chemistry. A metallic glass may show little difference. If both glasses agree with classical potentials, that is a useful negative result: it would justify decades of pair-potential yielding studies for real materials.

Where This Sits Relative to Prior Work

The cyclic-shear literature supplies the protocol, the observables, and the reference behavior. Oscillatory athermal quasistatic deformation of a model glass established the reversible-to-diffusive onset in Lennard-Jones mixtures. The yielding transition in amorphous solids under oscillatory shear deformation located yield near $\gamma_{max}\approx 0.07$–$0.08$ in the Kob-Andersen glass. Work on the threshold energy explains annealing-dependent ductile versus brittle yielding: The role of annealing in determining the yielding behavior of glasses under cyclic shear deformation and the landscape model in Models for the Yielding Behavior of Amorphous Solids. Correlation between plastic rearrangements and local structure in a cyclically driven glass tied plastic susceptibility to local structure: mobile atoms have high excess pair entropy and few tetrahedra, and icosahedra shield. All of these use pairwise or empirical three-body forces. We reuse their protocol and metrics unchanged and change only where the forces come from.

Three pieces of machine-learning work motivate and constrain the design. Predicting Properties of Amorphous Solids with Graph Network Potentials ran a universal DFT-trained equivariant network (NequIP) over about 28,800 melt-quenched binary glasses. It reported a zero-shot error of 38 meV/atom against about 10,000 DFT relaxations. That shows DFT-quality amorphous structures are affordable, but it validated only relaxed energies, never deformation. Predicting and Interpreting Energy Barriers of Metallic Glasses with Graph Neural Networks regressed activation barriers in Cu$_{64}$Zr$_{36}$ with a Pearson correlation of 0.7859, against 0.3614 for a local-sampling baseline. It found barriers are controlled by structure 5–10 Å away. This constrains us in two ways. The barrier labels came from classical potentials, and the model is static. Its medium-range finding also warns that a short-cutoff potential may miss what controls instabilities, so we test cutoff explicitly. De Novo Atomistic Discovery of Disordered Mechanical Metamaterials by Machine Learning showed that tuning bond directionality in a Stillinger-Weber potential strongly changes disordered mechanics. That is evidence that directional bonding, which DFT captures and pair laws do not, is a mechanical control variable. But the tuned parameters were never checked against real electronic structure, and the loading was never cyclic.

Research Program

We study two chemistries. Cu$_{64}$Zr$_{36}$ is a metallic glass with mature classical embedded-atom potentials and an existing barrier benchmark. Amorphous silicon is a covalent network where bond directionality dominates and the Stillinger-Weber potential is the classical standard. For each chemistry, the program has three stages.

The first stage is ground truth. We run athermal quasistatic cyclic shear directly with DFT (a PBE-type functional) on cells of roughly 200–300 atoms, for a few cycles at several amplitudes. Athermal quasistatic shear means small affine strain steps, each followed by energy minimization, at zero temperature and vanishing strain rate. At each step we record the stress, the energy, and the lowest nonzero Hessian eigenvalue $\lambda_{min}(\gamma)$. The Hessian eigenvalue measures how close the current minimum is to vanishing. A plastic event is a saddle-node instability, where $\lambda_{min}\propto(\gamma^*-\gamma)^{1/2}$ as the strain $\gamma$ approaches the instability strain $\gamma^*$. The square-root form is the generic behavior of a minimum merging with a saddle. So errors in the curvature near $\gamma^*$ translate directly into errors in where events happen. This small-cell dataset is the referee for everything that follows.

The second stage is event-targeted training. We start from a universal DFT-trained potential and fine-tune it with a fixed budget of new DFT labels. The labels come from configurations the potential flags as near instability, either because its own $\lambda_{min}$ is small or because an ensemble of models disagrees. Configurations are cut from large sheared cells as clusters or periodic sub-cells. The loss adds a term on low-lying Hessian eigenvalues, or equivalently on force derivatives along the softest mode, to the usual energy and force terms. This is the ingredient we isolate by ablation.

The third stage is the science run. With the validated potential, we cyclically shear cells of about $10^4$ atoms at many amplitudes and several annealing levels, for hundreds of cycles. We measure $\gamma_c$ from the stroboscopic diffusion coefficient, which scales as $D\sim(\gamma_{max}-\gamma_c)^\beta$. We also measure the threshold energy and the ductile-versus-brittle signature, and the avalanche and cluster size distributions for comparison with the 2D exponents in Yielding transition of a two dimensional glass former under athermal cyclic shear deformation. Separately, we take snapshots of mobile and immobile regions and compute DFT electronic diagnostics on them: local density of states, Bader charges, and, for CuZr, the Cu/Zr $d$-band center. We then ask whether any electronic descriptor predicts which atoms rearrange beyond what excess pair entropy and tetrahedrality already predict.

Baselines

The naive reference is the classical potential for each chemistry: embedded-atom for CuZr and Stillinger-Weber for silicon. Both are run through the identical shear protocol, cell sizes, annealing schedules, and cycle counts. This is what the field would do by default, and it is the result our DFT-level numbers must differ from to matter. We also report the Kob-Andersen yield range of 0.07–0.08 only as a qualitative anchor, not a target. For $\gamma_c$ or threshold energy in CuZr or amorphous silicon under cyclic shear, no reported result was found on Lacuna. We will produce these numbers ourselves for every model.

The strongest existing method is the zero-shot universal equivariant potential of Predicting Properties of Amorphous Solids with Graph Network Potentials, whose reported fidelity is 38 meV/atom MAE on relaxed amorphous structures. It is the strongest competitor because it already delivers DFT-quality amorphous energies at scale with no extra labels. If it reproduces small-cell DFT instability strains, our targeted training is unnecessary. It runs on the same cells and protocol, with no fine-tuning.

The key ablation uses the same architecture, cutoff, and DFT label budget as our method, with the fine-tuning labels drawn from random snapshots of sheared trajectories rather than near-instability configurations, and with no Hessian loss term. A second ablation refits a classical potential of the same functional form as the naive baseline on our full DFT dataset. This separates "more DFT data" from "many-body flexibility." A third ablation varies the network cutoff between 5 and 10 Å, prompted by the medium-range finding of Predicting and Interpreting Energy Barriers of Metallic Glasses with Graph Neural Networks.

Beating a baseline means lower error, on held-out small-cell DFT trajectories, in three quantities: instability strain $\gamma^*$, stress-drop size, and $\lambda_{min}$ within 1% strain of events. The margin must exceed the spread across independently trained seeds. Only potentials that pass this check are used to report $\gamma_c$.

Why not just…?

Why not just use the established classical potential? For CuZr this is a real threat. Embedded-atom potentials were fitted to metallic bonding and may place $\gamma_c$ correctly. We therefore treat CuZr partly as a control. The small-cell DFT comparison of instability strains settles it directly: if the embedded-atom potential matches DFT as well as our potential does, we report that classical yielding results transfer to this metallic glass. For silicon, the metamaterial study's sensitivity to three-body bending parameters suggests the fixed Stillinger-Weber form is less trustworthy, but that is an expectation, and the same check decides it.

Why not just run the universal DFT-trained potential zero-shot? Its reported accuracy is on relaxed melt-quenched structures. Nothing in the file tests it on configurations near a vanishing minimum, where it has no training data. This is the most likely point of failure. It is also the cheapest alternative, so it is our main baseline. If its instability-strain errors match those of the fine-tuned model, we drop the targeted-training claim and keep only the physics results.

Why not just add more random DFT data? More data does help, but sheared trajectories spend almost all their steps in elastic segments, far from instabilities. Random sampling therefore spends most of the label budget where the potential is already accurate. The equal-budget random-snapshot ablation tests exactly this. If it matches targeted sampling, the targeted-sampling contribution disappears.

Why not run classical shear and apply DFT only afterwards to snapshots? Post-hoc electronic analysis can describe the electrons in configurations a classical potential chose. It cannot say whether the electrons would have chosen different configurations, and $\gamma_c$ and the threshold energy are exactly such choices. We do use post-hoc DFT for the descriptor study, but only on trajectories generated with the validated potential.

Decisive Checks

Three results decide the proposal. First, on held-out direct-DFT cyclic shear of small cells, does targeted fine-tuning reduce errors in $\gamma^*$ and $\lambda_{min}$ relative to the zero-shot universal potential and the random-label ablation? If not, the methodological claim fails. Second, at $10^4$ atoms, does $\gamma_c$ or the threshold energy from the validated potential differ from the classical baseline by more than the spread across seeds and samples? We expect a clear gap in amorphous silicon and at most a small one in CuZr. Agreement in both is reported as a negative result. Third, does an electronic descriptor from snapshot DFT improve prediction of which atoms rearrange, measured by area under the ROC curve over excess pair entropy plus tetrahedrality? This determines whether the electronic view adds information about where plasticity starts, not only about where yield occurs.