Geometric Encoding of Cyclic Mechanical Fatigue in Model Glasses: Graph Neural Networks as Structural Probes for Battery-Relevant Degradation
Year: 2026
Type: preprint
Abstract: Solid-state battery glass electrolytes degrade under repeated charge and discharge cycling, yet the atomic-scale structural changes underlying this fatigue remain poorly characterised. Here we investigate whether cyclic mechanical fatigue—simulated as oscillatory volumetric strain applied to a Lennard-Jones model glass—encodes a learnable geometric signature detectable by graph neural networks (GNNs). Using a GATv2 architecture and a systematic ablation study, we find that classification between pristine (cycle 0) and heavily fatigued (cycles 300–400) glasses reaches 91.67 ± 4.22% accuracy using 8-dimensional bond-length statistics as node features, rising to 98.33 ± 1.49% (AUC = 0.9975 ± 0.0022) under per-sample instance normalisation. The improvement under normalisation—which explicitly removes all global magnitude information—establishes that the discriminative signal resides in the relational geometry of local bond networks, not in trivial density or volume changes. Ablating distribution-shape features (skewness, quartiles) reduces classification accuracy by 5.3 percentage points and collapses cycle-number regression to near-random performance (R² = 0.006), confirming that asymmetric bond-length damage is the dominant structural fingerprint of fatigue. Continuous degradation tracking achieves R² = 0.311 ± 0.039 under instance normalisation. Permutation importance analysis identifies a minimal sufficient descriptor set of four features {r_max, r̄, σ_r, γ̃} that exactly preserves classification accuracy (91.67 ± 2.36%) while halving feature dimensionality and improving regression R² from 0.062 to 0.264 without normalisation—demonstrating that the dropped features were actively degrading convergence rather than contributing signal. Per-atom energy validation reveals a striking hierarchy inversion: the dominant classification feature (r_max, ΔAUC = 0.601) correlates negligibly with local potential energy (Spearman ρ = −0.093), while the strongest energy proxy (r_min, ρ = −0.812) ranks seventh in fatigue discrimination—establishing that the GATv2 probe detects irreversible structural history encoded in stretched-bond geometry rather than the instantaneous energetic state. A Kovacs memory effect experiment further validates the 8D feature vector as a genuine structural probe: bond-length standard deviation and skewness track the NVT Kovacs relaxation with the same timescale as macroscopic energy recovery, while mean non-affine displacement continues to grow linearly after bulk equilibration, confirming that the GNN features occupy the physically privileged observational space between thermodynamic bulk observables and the full microscopic configuration. Attention-weight analysis reveals that fatigue manifests as isolated, point-like single-atom defects with a flat spatial correlation length across all 400 cycles (ξ ≈ 0.05σ), providing geometric evidence against Shear Transformation Zone-mediated damage and in favour of diffuse plastic rearrangement; Jaccard hotspot persistence (J ≈ 0.08–0.12) confirms that the damaged-atom population rotates continuously rather than accumulating at static structural weak points. Finite-size scaling at N = 1024 validates all qualitative findings: instance-normalised classification reaches 100.00 ± 0.00% (AUC = 1.000) and regression improves to R² = 0.428 ± 0.041, with normalisation delivering approximately 6× faster convergence and complete elimination of fold instabilities relative to the unnormalised baseline. These findings extend the geometric-encoding framework established for thermal history to mechanical cycling history, and provide a methodological and physical foundation for applying GNN-based structural probes to non-destructive fatigue diagnostics in experimentally characterised solid electrolyte systems.
Source: ChemRxiv
Author: Saurab Singh
Language: English
Subfield: Materials Chemistry
Field: Materials Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7135201610