Interpretability of linear regression models of glassy dynamics
Anand Sharma, Chen Liu, Misaki Ozawa, Daniele Coslovich · 2025
Read Interpretability of linear regression models of glassy dynamics by Anand Sharma, Chen Liu, Misaki Ozawa, Daniele Coslovich (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2508-15933.txt.
Introduction
Anand Sharma, Chen Liu, Misaki Ozawa, and Daniele Coslovich ask whether simple models of glassy motion can still teach physics, in a 2025 paper. Data-driven models can predict how a glass-forming liquid will move from a snapshot of its structure. But a model that predicts well may still hide what drives the motion, if its inputs overlap too much.
Working with a flat model glass, the team find that many standard structure measures do overlap, which blurs what each one adds. A regularized fit trims some blur but stays too tangled to read. Only after shrinking the input set do they get models that are both accurate and readable. Those models point to local packing and mix shifts as the key drivers in this glass. The study is a caution for anyone reading physics off a fitted model.
Catalog record
| CREATOR | Anand Sharma, Chen Liu, Misaki Ozawa, Daniele Coslovich |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 101 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2508-15933
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Interpretability of linear regression models of glassy dynamics Anand Sharma, Chen Liu, Misaki Ozawa, Daniele Coslovich Abstract Data-driven models can accurately describe and predict the dynamical properties of glass-forming liquids from structural data. Accurate predictions, however, do not guarantee an understanding of the underlying physical phenomena and the key factors that control them. In this paper, we illustrate the merits and limitations of linear regression models of glassy dynamics built on high-dimensional structural descriptors. By analyzing data for a two-dimensional glass model, we show that several descriptors commonly used in glass‑transition studies display multicollinearity, which hinders the interpretability of linear models. Ridge regression suppresses some of the shortcomings of multicollinearity, but its solutions are not succinct enough to be physically interpretable. Only by using dimensional reduction techniques we eventually obtain linear models that strike a balance between prediction accuracy and interpretability. Our analysis points to a key role of local packing and composition fluctuations in the glass model under study. I Introduction As the temperature of glass-forming liquids decreases, the relaxation timescale increases dramatically, while structural changes remain relatively modest Berthier and Biroli 2011 ; Binder and Kob 2011 ; Ediger et al. 1996 . Moreover, the dynamics exhibits pronounced spatial fluctuations, characterized by regions of high and low mobility, referred to as dynamic heterogeneities Ediger 2000 ; Berthier et al. 2011 ; Karmakar et al. 2014 . Despite the vivid patterns of dynamic heterogeneities, however, static snapshots of the system appear homogeneous and lack distinct features, at least to the naked eye. This has been confirmed through direct observations in computer simulations and colloidal glass experiments Widmer-Cooper et al. 2004 ; Weeks 2017 . To identify subtle but significant structural changes linked to dynamics, various structural order parameters based on physical intuition have been investigated Coslovich and Pastore 2007 ; Patrick Royall et al. 2008 ; Widmer-Cooper et al. 2008 ; Hocky et al. 2014 ; Jack et al. 2014 ; Turci et al. 2017 ; Royall and Williams 2015 ; Tanaka et al. 2019 ; Kapteijns et al. 2021 ; Sharma et al. 2022 . Similar efforts have been done to identify the structural origins of plastic events in amorphous solids under loading Manning and Liu 2011 ; Richard et al. 2020 ; Baggioli et al. 2021 ; Wu et al. 2023 . Recent advances in machine learning have demonstrated that glassy dynamics, including dynamic heterogeneities, can be accurately described and predicted from local structural information Cubuk et al. 2015 ; Jung et al. 2025 . A range of machine learning techniques has been applied to this problem, including support vector machines Cubuk et al. 2015 ; Schoenholz et al. 2016 , multi-layer perceptrons Jung et al. 2023 , and graph neural networks Bapst et al. 2020 ; Shiba et al. 2023 ; Jiang et al. 2023 ; Pezzicoli et al. 2024 . This research has primarily progressed in two directions. The first involves increasing the complexity of machine learning architectures, employing deep neural networks with a large number of parameters and taking advantage of cutting-edge methods Bapst et al. 2020 ; Shiba et al. 2023 ; Pezzicoli et al. 2024 . The second direction focuses on integrating domain knowledge from physics in data-driven models Boattini et al. 2021 ; Jung et al. 2023 . These physics-informed approaches enable high predictive accuracy while maintaining relatively simple model architectures. However, achieving accurate predictions alone does not guarantee an understanding of the underlying mechanisms driving the phenomenon under investigation Teney et al. 2022 ; Swain et al. 2024 . From the perspective of fundamental physics research, it is crucial that data-driven models provide physically interpretable results, which must be robust and expressed in a succinct form. A growing body of glass transition studies explores the issue of interpretability in deep‑learning models or non-linear models, using a range of approaches Swanson et al. 2020 ; Font-Clos et al. 2022 ; Oyama et al. 2023 ; Ciarella et al. 2023 ; Janzen et al. 2023 ; Janzen et al. 2024 ; Liu et al. 2024 ; Jung et al. 2024 ; Wang et al. 2024 ; Yoshikawa et al. 2025 ; Wang et al. 2020 ; Liu et al. 2023 . While these works offer useful clues, the underlying deep networks remain highly complex and only partially transparent. Consequently, there is still a clear need for explicitly interpretable models whose solutions admit a direct physical reading. Remarkably, recent studies have demonstrated that simple linear models, when combined with domain knowledge of glassy dynamics, such as coarse-graining techniques, can describe dynamic heterogeneities with remarkable accuracy Boattini et al. 2021 ; Alkemade et al. 2022 . In some cases, their accuracy is comparable to that of more complex deep learning models. This outcome is particularly desirable, because it is often believed that simpler models not only provide accurate predictions at a minimal computational cost, but can also offer greater interpretability Rudin 2019 . We caution, however, that employing a linear model per se is not sufficient to extract meaningful physical information: in high‑dimensional settings, linear models are often plagued by numerical instabilities, due to the so-called multicollinearity Montgomery et al. 2021 , that can obscure interpretation. Moreover, their solution may still be too high-dimensional to convey a physical meaning. The aim of this study is to address these issues in linear models for glassy dynamics and introduce remedies that restore interpretability. To tackle the issue of interpretability in a rigorous manner, we will study quantitatively the consequences of multicollinearity in linear regression models of glassy dynamics, using a two-dimensional glass-forming liquid model. First, we will demonstrate how multicollinearity leads to instability in weight estimation within linear regression, thereby hindering the interpretation of feature importance. These problems affect several structural descriptors used in recent glass transition studies. Next, we will explore strategies to mitigate the effects of multicollinearity and to identify a set of low-dimensional linear models that achieve good performance accuracy with a minimum of structural information. We will critically evaluate and discuss the advantages and limitations of each approach, providing a comprehensive assessment of linear models of the structure-dynamics relationship in a glass-forming liquid model. The paper is organized as follows. Section II defines the problem under investigation. Section III describes the physical model and our structure-dynamics dataset. Sections IV and Section V introduce simple linear regression models of glassy dynamics and demonstrate the effects of multicollinearity on the instability of the estimated weights. Section VI examines feature selection and extraction approaches that cope with multicollinearity while also reducing the dimensionality of the problem. Sections VII and VIII offer a critical outlook on our results and summarize the key findings of the work. II Problem statement In data-driven modeling of glassy dynamics, the problem is typically formulated as follows Jung et al. 2025 . Several variables, called features, are computed in order to characterize the local structural environment around each particle. These structural features are collected into a descriptor that provides a high-dimensional representation of the local structure. Linear regression models estimate the dynamical observable {\bf Y} (in this study, the dynamic propensity Widmer-Cooper and Harrowell 2006 ) as a linear combination of M input structural features, {\bf X}^{(1)},{\bf X}^{(2)},\ldots,{\bf X}^{(M)} , using