Abstract: Rational design of transition metal complexes with desired optical properties is a major challenge due to high computational costs of quantum-chemical methods that can deliver quantitatively reliable results. We present a machine learning framework for predicting absorption and emission maxima in both transition metal coordination compounds and organic chromophores using joint training on a combined experimental data set. Our featurization strategy integrates ligand environment fingerprints (Morgan), metal center features (Coulomb matrices), and topological descriptors from persistent homology analysis. The combined training data set comprises 19,733 absorption and 2675 emission measurements for 17,359 metal complexes (with focus on Ir, Rh, Pt, and Ru systems) and 17,294 absorption and 18,141 emission measurements for 7065 organic molecules across 365 solvents. Among several architectures evaluated, multilayer perceptrons provide the best absorption predictions (RMSE = 33.5 nm, R2 = 0.83, Pearson r = 0.92 for metal–organic compounds), while gated recurrent units are optimal for emission (RMSE = 41.7 nm, R2 = 0.83, Pearson r = 0.90). Models trained jointly on both data sets show good universal applicability with moderate accuracy trade-offs: RMSE increases by approximately 7–19 nm for organic compounds compared to specialized models, and for metal–organic compounds, RMSE increases by 1–2 nm. In contrast, models trained on organic data alone fail catastrophically when applied to metal complexes (R2 = 0.01). For a test set of 35 metal complexes including metal centers beyond the main training distribution (V, W, Cu, and Os in addition to Ir, Rh, Pt, and Ru), our best models achieve an RMSE of ∼28 nm for absorption maxima, comparable to TDDFT-O3LYP predictions but at substantially lower computational costs. SHAP analysis reveals that Coulomb matrix descriptors are most important for metal complex predictions, while Morgan fingerprints prevail for purely organic compounds. The presented approach enables efficient screening of candidate compounds for various photophysical applications orders of magnitude faster than TDDFT calculations.

Cite: Ilin E.A. , Ilina V.V. , Ioffe I.N. , Medvedev M.G. , Goryunkov A.A.
Machine Learning Prediction of Absorption and Emission: A Unified Approach for Metal–Organic and Organic Chromophores
Journal of Physical Chemistry A. 2026. V.130. N24. P.4670-4682. DOI: 10.1021/acs.jpca.6c00682 WOS Scopus OpenAlex

Dates:

Submitted:	Jan 31, 2026
Published online:	Jun 18, 2026

Identifiers:

≡ Web of science:	WOS:001788420900001
≡ Scopus:	2-s2.0-105042096990
≡ OpenAlex:	W7163988928

Altmetrics: