Graph Neural Networks for Accelerated Prediction of Excited-State Properties in Transition-Metal Complexes
- 1Department of Chemistry
- 2University of Alberta
- 3Alberta Machine Intelligence Institute
Abstract
Machine learning (ML) has been increasingly applied to the ground-state properties of transition-metal complexes (TMCs), whereas its use for predicting their stateresolved photophysical properties remains comparatively limited. Here, we evaluate nine ML configurations: a descriptor-based XGBoost model and eight graph neural network (GNN) configurations, using excited-state reference data from tmQMg* together with molecular, structural, and electronic information from the tmQM dataset series. The models were trained to predict the 30 lowest vertical singlet excitations, including excitation energies (∆E), oscillator strengths ( f ), absorption properties, charge-transfer (CT) character, and the metal contributions to the occupied and virtual natural transition orbitals (NTOs). GNNs substantially outperformed XGBoost for ∆E. Among the investigated architectures, the polarizable atom interaction neural network (PaiNN) provided the most favorable balance between accuracy and input requirements, reproducing ∆E with a mean absolute error of 0.124 eV while requiring only an XYZ file and formal molecular charge. Furthermore, PaiNN reliably captured excited-state character, resulting in 87.0% accuracy for CT classification and occupied-and virtual-NTO metal-fraction MAEs of 0.082 and 0.058, respectively, indicating that it generally distinguishes metal-and ligand-centered orbital character well. Individual f values were more difficult to predict accurately; however, combining the predicted ∆E and intensities reproduced the principal features of the broadened absorption spectra. Overall, these results establish PaiNN as a fast and reliable predictor of multiple TMC excited-state properties from minimal molecular input, providing a practical route for large-scale photophysical screening and substantially reducing the need for time-dependent density functional theory (TD-DFT) calculations and subsequent excited-state analysis.
Related
- Alginate/chitosan polyelectrolyte complexes: A comparative study of the influence of the drying step on physicochemical properties
- Building the Data Foundations for the Design and Discovery of Transition Metal Complexes
- Diiron(I) Complexes with Promising Anticancer Properties Obtained by CO-To-Isocyanide Substitution of [Fe2Cp2(CO)4]
- Structure-activity relationship of Cu(II) and Pd(II) thiosemicarbazone complexes with anticancer and antibacterial properties
- Mononuclear lanthanide( iii )-based complexes with β-thiooxoketone and β-diketone as ligands: effects of soft (S) vs. hard (O) donor substitution on magnetic and luminescent properties
BibTeX
@misc{elayan2026graph,
title = {{Graph Neural Networks for Accelerated Prediction of Excited-State Properties in Transition-Metal Complexes}},
author = {Ismael A. Elayan and J. Diego Castro-Miyashiro and J. Terence Blaskovits},
howpublished = {ChemRxiv},
year = {2026},
month = sep,
doi = {10.26434/chemrxiv.15008673/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15008673/v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - Graph Neural Networks for Accelerated Prediction of Excited-State Properties in Transition-Metal Complexes
AU - Ismael A. Elayan
AU - J. Diego Castro-Miyashiro
AU - J. Terence Blaskovits
PY - 2026
DA - 2026/09/10
PB - ChemRxiv
DO - 10.26434/chemrxiv.15008673/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15008673/v1
AB - Machine learning (ML) has been increasingly applied to the ground-state properties of transition-metal complexes (TMCs), whereas its use for predicting their stateresolved photophysical properties remains comparatively limited. Here, we evaluate nine ML configurations: a descriptor-based XGBoost model and eight graph neural network (GNN) configurations, using excited-state reference data from tmQMg* together with molecular, structural, and electronic information from the tmQM dataset series. The models were trained to predict the 30 lowest vertical singlet excitations, including excitation energies (∆E), oscillator strengths ( f ), absorption properties, charge-transfer (CT) character, and the metal contributions to the occupied and virtual natural transition orbitals (NTOs). GNNs substantially outperformed XGBoost for ∆E. Among the investigated architectures, the polarizable atom interaction neural network (PaiNN) provided the most favorable balance between accuracy and input requirements, reproducing ∆E with a mean absolute error of 0.124 eV while requiring only an XYZ file and formal molecular charge. Furthermore, PaiNN reliably captured excited-state character, resulting in 87.0% accuracy for CT classification and occupied-and virtual-NTO metal-fraction MAEs of 0.082 and 0.058, respectively, indicating that it generally distinguishes metal-and ligand-centered orbital character well. Individual f values were more difficult to predict accurately; however, combining the predicted ∆E and intensities reproduced the principal features of the broadened absorption spectra. Overall, these results establish PaiNN as a fast and reliable predictor of multiple TMC excited-state properties from minimal molecular input, providing a practical route for large-scale photophysical screening and substantially reducing the need for time-dependent density functional theory (TD-DFT) calculations and subsequent excited-state analysis.
ER -