Beyond overcomplication: a linear model suffices to decode hidden structure-property relationships in glasses

Establishing reliable and interpretable structure-property relationships in glasses is a longstanding challenge in condensed matter physics. While modern data-driven machine learning techniques have proven highly effective in establishing structure-property correlations, many models are criticized for lacking physical interpretability and being task-specific. In this work, we identify an approximate linear relation between structure profiles and disorder-induced responses of glass properties based on first order perturbation theory. We analytically demonstrate that this relationship holds universally across glassy systems with varying dimensions and distinct interaction types. This robust theoretical relationship motivates the adoption of linear machine learning models, which we show numerically to achieve surprisingly high predictive accuracy for structure-property mapping in a wide variety of glassy materials. We further devise regularization analysis to further enhance the interpretability of our model, bridging the gap between predictive performance and physical insight. Overall, this linear relation establishes a simple yet powerful connection between structural disorder and spectral properties in glasses, opening a new avenue for advancing their studies.

💡 Research Summary

The paper tackles the longstanding problem of establishing reliable, interpretable structure‑property relationships (SPRs) in glassy materials. While modern machine‑learning (ML) approaches, especially deep neural networks, have achieved impressive predictive performance, they are often criticized for being black‑box models that provide little physical insight and for requiring large training datasets. The authors propose a fundamentally different strategy: they derive, from first‑order perturbation theory, an approximate linear relationship between a simple two‑body structural descriptor—the radial distribution function (RDF) g(r)—and disorder‑induced changes in a broad class of phonon‑related properties such as the phonon density of states (PDOS), Raman spectra, and inelastic neutron or X‑ray scattering spectra.

The theoretical development begins by treating the atomic positions in a glass as small displacements Δr_i from a reference crystalline configuration. Expanding the RDF to first order yields a set of linear equations linking g(r) to the displacement field. The authors then consider the Hessian matrix H, whose eigenvalues λ_n determine the vibrational frequencies ω_n = √λ_n. By expressing the perturbation of H as ΔH_ij = Σ_k Φ_ijk Δr_k, where Φ_ijk are third‑order force constants, they show that the eigenvalue shifts Δλ_n are linear functions of the same displacements. To guarantee the validity of the linear approximation, they invoke Weyl’s inequality and results from random matrix theory for Gaussian‑distributed Δr_i, establishing that the operator norm of ΔH is bounded with high probability, and consequently |Δλ_n| remains small. This rigorous analysis demonstrates that, irrespective of the specific interatomic potential or spatial dimensionality, the mapping from RDF to phonon‑related observables is essentially linear for modest disorder.

Guided by this insight, the authors construct a universal linear regression model: y = Wθ g + bθ, where g is a vectorized RDF (obtained by discretizing r into bins), y is the target property vector (e.g., a discretized PDOS), and Wθ, bθ are trainable weights and bias shared across all structures. The model is deliberately parsimonious—its parameter count is orders of magnitude smaller than that of typical convolutional neural networks (CNNs) used in prior work such as SPRamNet.

The model’s performance is evaluated on four diverse glassy systems: (i) two‑dimensional amorphous monolayer carbon (AMC), (ii) periodic Lennard‑Jones glasses, (iii) bulk amorphous silicon carbide (SiC), and (iv) a ternary amorphous CuAlZr high‑entropy alloy. For each system, the authors generate a wide spectrum of disorder by varying quench rates, defect concentrations, or composition, compute RDFs and reference PDOS via molecular dynamics and lattice‑dynamics calculations, and train the linear model on a subset of the data. Predictive accuracy is quantified using root‑mean‑square error (RMSE) and coefficient of determination (R²). Across all cases the linear model achieves R² ≥ 0.96 and RMSE reductions of 10–20 % relative to state‑of‑the‑art CNN baselines, despite using far fewer training samples. Notably, the model captures subtle disorder‑induced changes in high‑frequency PDOS features (≈65 THz) that are most sensitive to bond‑stretching vibrations.

Interpretability is a central theme. By applying Elastic Net regularization (a combination of L1 and L2 penalties) the authors induce sparsity in Wθ, allowing them to identify which radial bins of the RDF contribute most strongly to each property. The analysis reveals that short‑range correlations (r ≈ 1.3–1.5 Å) dominate high‑frequency vibrational responses, while medium‑range features influence lower‑frequency acoustic modes—findings that align with established physical intuition. Regularization also mitigates over‑fitting, ensuring robust performance even when the training set is limited.

In conclusion, the study demonstrates that a simple linear ML model, grounded in perturbation theory, can rival or surpass sophisticated nonlinear architectures for predicting glassy material properties. The approach offers three key advantages: (1) computational efficiency and minimal data requirements, (2) direct physical interpretability through the RDF‑to‑property mapping, and (3) universal applicability across dimensions, interaction types, and compositional complexity. The authors suggest future extensions to electronic and thermal transport properties, as well as hybrid models that incorporate modest nonlinear corrections while retaining interpretability. This work thus opens a pragmatic pathway for rapid, insight‑driven materials design in disordered systems.