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

Edward Danquah Donkor, Alessandro Laio, Ali Hassanali

Abstract

Machine-learning (ML) has become a key workhorse in molecular simulations. Building an ML model in this context involves encoding the information of 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 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 entering in the SOAP definition in encoding chemical information.

Introduction

Machine learning (ML) has revolutionized molecular simulations, primarily through the use of "atomic descriptors" that encode local 3D atomic environments into high-dimensional mathematical vectors. Despite their success, these models often function as "black boxes," leaving researchers uncertain about the specific physical or chemical principles being captured.

Comparison between SOAP descriptors and traditional chemical order parameters used to characterize liquid water structure.

For decades, physical chemists have relied on intuitive "order parameters"—such as hydrogen bonding, local density, or tetrahedrality—to describe liquid water. This study addresses a fundamental interpretability gap: Do modern ML descriptors, specifically the Smooth Overlap of Atomic Positions (SOAP), encode the same structural information as traditional order parameters? Furthermore, the research investigates how this relationship evolves when moving from bulk water to the hydrophobic air-water interface.

Methodology

The study utilized Molecular Dynamics (MD) simulations of liquid water using the TIP4P/2005 and TIP3P rigid water models in LAMMPS. The dataset comprised approximately 5,000 local water environments sampled every 4 picoseconds. The Willard-Chandler Interface (WCI) method was employed to distinguish between surface and bulk molecules.

The researchers computed the SOAP power spectrum for these environments and compared them against a suite of standard order parameters, including Orientational and Translational Tetrahedral Order ($q_{tet}$, $S_k$), Local Structural Index (LSI), distance to the 5th oxygen ($d_5$), Coordination Number (C.N.), Number of Hydrogen Bonds ($N_{H.B.}$), Voronoi Density ($\rho_{voro}$), and SOAP distance from ice ($d_{ICE}$).

To bridge these disparate metrics, the authors applied the Information Imbalance (IB) statistical technique. IB measures the relative information content between two distance metrics on the same data. It essentially asks: if two environments are neighbors in SOAP space, are they also neighbors in the space defined by a specific order parameter? An IB value near 0 indicates perfect predictability, while a value near 1 suggests no correlation.

Results and Discussion

Non-Equivalence of ML and Chemical Intuition

The primary finding is that ML descriptors and traditional order parameters are fundamentally not equivalent. Environments deemed nearly identical by SOAP can appear significantly different under traditional metrics, and vice versa. This discrepancy largely stems from the nature of the metrics: traditional parameters often rely on rigid geometric cutoffs (e.g., a 30° angle for hydrogen bonds), leading to topological jumps. In contrast, SOAP uses smooth Gaussian densities, treating small geometric shifts as nearly identical.

The Role of Hyperparameters

The SOAP $\sigma$ hyperparameter, which controls Gaussian smearing, plays a critical role. While a default $\sigma = 1.0$ Å is common, reducing it to $0.25$ Å allows SOAP to capture finer chemical details. However, this increased resolution makes it harder to explain SOAP similarity using traditional parameters; it requires a combination of at least five order parameters to achieve the same predictive power that three parameters provide for the coarser $\sigma = 1.0$ Å version.

Predictability and Independence

SOAP effectively predicts Coordination Number and Voronoi density using only a small fraction of its 4410-dimensional vector. However, it struggles to predict the Number of Hydrogen Bonds, LSI, and $q_{tet}$. Interestingly, the traditional order parameters themselves were found to be largely independent; for instance, knowing a molecule's tetrahedrality provides almost no information about its local density.

Surface vs. Bulk Dynamics

The study also found that structural deviations at the air-water interface are highly localized. Information Imbalance metrics indicate that water molecules just 3.0 Å from the surface are virtually indistinguishable from bulk water in terms of descriptor behavior.

Conclusion

This work highlights a significant tension between the smooth, generic nature of ML features and the strict geometric definitions of traditional chemical metrics. While SOAP descriptors are powerful, they may lack the specific resolution required to capture topological features like hydrogen bond networks perfectly. Future developments in ML models for molecular systems might benefit from augmenting generic descriptors with features that directly probe chemical fluctuations, or conversely, by refining traditional metrics to be less sensitive to arbitrary geometric cutoffs.

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 of 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 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 viceversa, 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 entering 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