Do Machine-Learning Atomic Descriptors and Order Parameters Tell the Same Story? The Case of Liquid Water

Edward Danquah Donkor, Alessandro Laio, Ali Hassanali

TL;DR

This paper investigates whether popular machine-learning atomic descriptors (like SOAP) and traditional, chemically-inspired order parameters (like hydrogen bonding or tetrahedrality) capture the same structural information in liquid water. Using a statistical test called Information Imbalance, the authors reveal that these two descriptions are fundamentally not equivalent, highlighting the tension between the smooth, generic nature of ML features and the strict geometric cutoffs of traditional chemical metrics.

What Problem Does This Paper Address?

Machine learning (ML) has become a massive part of molecular simulations, heavily relying on "atomic descriptors" to encode the local 3D environment of atoms into mathematical vectors. However, ML models are notoriously black boxes. It is often unclear what physical or chemical principles these descriptors are actually capturing.

For decades, physical chemists have studied liquid water using intuitive "order parameters"—metrics like the number of hydrogen bonds, local density, or how perfectly tetrahedral a water molecule's surroundings are. This paper asks a fundamental interpretability question: Do modern ML descriptors (specifically the Smooth Overlap of Atomic Positions, or SOAP) encode the same structural story as the traditional order parameters we’ve relied on for years? Furthermore, how does this relationship change between bulk water and the hydrophobic air-water interface?

Methods

  • Molecular Dynamics (MD) Simulations: The authors simulated liquid water using the TIP4P/2005 (and TIP3P) rigid water models in LAMMPS, modeling both bulk water and an air-water interface.
  • Descriptors & Order Parameters: They computed the SOAP power spectrum for local water environments. They compared this against a suite of standard order parameters: Orientational/Translational Tetrahedral Order ($q_{tet}$, $S_k$), Local Structural Index (LSI), distance to the 5th oxygen ($d_5$), Coordination Number, Number of Hydrogen Bonds, Voronoi Density, and SOAP distance from ice.
  • Information Imbalance (IB): To compare these vastly different metrics, they used a recently developed statistical technique called Information Imbalance. IB measures the relative information content between two different distance metrics on the same data. Essentially, it asks: If two water environments are nearest neighbors in SOAP space, are they also nearest neighbors in Hydrogen-Bond space? (An IB of ~0 means perfect predictability; ~1 means no correlation).

Datasets

The dataset consists of approximately 5,000 local water environments sampled in-house from their MD simulations. Environments were chosen by randomly selecting a water molecule every 4 picoseconds. The Willard-Chandler Interface (WCI) method was used to categorize molecules into distinct layers (Surface vs. Bulk).

Key Insights and Ideas

  • ML and Chemical Intuition are Not Equivalent: Environments that are nearly identical according to SOAP can look completely different according to traditional order parameters, and vice versa.
  • The Problem with Rigid Cutoffs: Traditional parameters often rely on strict geometric cutoffs. For example, a standard hydrogen bond definition requires an angle of less than 30°. If an angle shifts from 29° to 31°, the traditional metric registers a lost hydrogen bond—a massive topological jump. SOAP, which uses smooth Gaussian densities, sees these two states as nearly identical.
  • The Importance of the $\sigma$ Hyperparameter: In SOAP, $\sigma$ controls the Gaussian smearing (blurriness) of the atomic positions. The default is often 1.0 Å. The authors found that reducing $\sigma$ to 0.25 Å drastically improves SOAP's ability to capture fine chemical details, though it requires a more complex combination of order parameters to explain.
  • Order Parameters are Independent: Surprisingly, the traditional order parameters themselves contain very little information about each other. Knowing a water molecule's tetrahedrality ($q_{tet}$) tells you almost nothing about its local density.

Key Results

  • Predicting SOAP: Even the best combination of three traditional order parameters can only predict standard SOAP ($\sigma=1.0$ Å) similarity with an Information Imbalance of ~0.45 (far from perfect). When $\sigma$ is reduced to 0.25 Å, it takes at least five order parameters to achieve the same predictive power.
  • Predicting Order Parameters: SOAP easily predicts Coordination Number and Voronoi density (requiring only 10-20 components of the 4410-dimensional SOAP vector). However, SOAP struggles significantly to predict the Number of Hydrogen Bonds, LSI, and $q_{tet}$.
  • Surface vs. Bulk: The structural deviations caused by the air-water interface are highly localized. The IB metrics show that water just ~3.0 Å away from the surface is virtually indistinguishable from bulk water in terms of how these descriptors behave.

Limitations and Future Work

The authors acknowledge that SOAP descriptors, designed to be generic, may lack the specific chemical resolution needed to perfectly capture topological features like the hydrogen bond network. To bridge this gap, they suggest future ML models could augment SOAP with features that directly probe hydrogen bond fluctuations. Conversely, they point out that traditional order parameters might be too sensitive to arbitrary geometric cutoffs, suggesting that the field should perform deeper sensitivity analyses on the geometric criteria used in chemistry-inspired metrics.

Summary updated

Abstract

Machine-learning (ML) has become a key workhorse in molecular simulations. Building an ML model in this context involves encoding the information on chemical environments using local atomic descriptors. In this work, we focus on the Smooth Overlap of Atomic Positions (SOAP) and their application in studying the properties of liquid water both in the bulk and at the hydrophobic air-water interface. By using a statistical test aimed at assessing the relative information content of different distance measures defined on the same data space, we investigate if these descriptors provide the same information as some of the common order parameters that are used to characterize local water structure such as hydrogen bonding, density, or tetrahedrality to name a few. Our analysis suggests that the ML description and the standard order parameters of the local water structure are not equivalent. In particular, a combination of these order parameters probing local water environments can predict SOAP similarity only approximately, and vice versa, the environments that are similar according to SOAP are not necessarily similar according to the standard order parameters. We also elucidate the role of some of the metaparameters in the SOAP definition in encoding chemical information.
BibTeX
@article{donkor2023do,
  author = {Edward Danquah Donkor and Alessandro Laio and Ali Hassanali},
  title = {{Do Machine-Learning Atomic Descriptors and Order Parameters Tell the Same Story? The Case of Liquid Water}},
  journal = {Journal of Chemical Theory and Computation},
  year = {2023},
  doi = {10.1021/acs.jctc.2c01205},
}
Figure