Deciphering the Solvation Structure of Aqueous ZnCl2 Solutions from X-ray Absorption Spectra Using the Interpretable Graph Neural Network | JoVE Visualize

Area of Science:

Computational Chemistry
Materials Science
Spectroscopy

Background:

Machine learning (ML) offers powerful predictive capabilities for spectroscopic observables from atomic structures.
Interpreting ML model predictions in terms of physical and chemical principles is crucial for broader scientific impact.
Predicting X-ray absorption spectra (XAS) for complex systems like concentrated solutions remains challenging for conventional methods.

Purpose of the Study:

To develop and validate a physics-guided graph neural network (GNN) model for predicting Zn K-edge X-ray spectroscopy (XAS) spectra.
To demonstrate the model's ability to interpret ML predictions in terms of physical and chemical principles.
To establish a general paradigm for interpretable ML in linking atomic structure, electronic structure, and spectroscopic observables.

Main Methods:

A physics-guided graph neural network (GNN) model was developed.
Training data were generated using ab initio XAS calculations on molecular dynamics snapshots from a ML interatomic potential.
Gradient-based attribution analysis was employed to interpret the GNN model's predictions.

Main Results:

The GNN model accurately reproduced experimental XAS spectra for aqueous ZnCl2 solutions across a wide range of concentrations (0.1 m to 30 m).
The model efficiently scaled to large, disordered liquid systems beyond the scope of traditional ab initio methods.
Attribution analysis revealed physically meaningful structure-spectrum relationships, including ligand-specific attributions reflecting orbital hybridization and bond-length attributions consistent with multiple-scattering theory.

Conclusions:

The developed GNN provides accurate and interpretable predictions of XAS spectra for complex liquid systems.
This work bridges data-driven prediction with electronic-structure theory, offering a new paradigm for interpretable ML in chemistry and materials science.
The interpretable ML approach successfully links atomic structure, electronic structure, and spectroscopic observables, advancing the understanding of chemical systems.