Skip to content
PaperFren

Neural nets predict TM spin states and bonds

Open paper intelligence

Graph-based ANNs trained on 2690 DFT geometries of Cr–Ni octahedral complexes predict spin splitting and metal–ligand distances without 3D coordinates.

Source

Predicting electronic structure properties of transition metal complexes with neural networks

Janet JP, Kulik HJ · Chemical science · 2017

doi.org/10.1039/c7sc01247kRead the full paper ↗114 citationscc by

Study at a glance

Design
Computational / modelling — ANN trained on DFT octahedral Cr–Ni complexes to predict spin states and bond lengths
N
N=2690 · 2,690 DFT geometry optimizations after filtering spin-contaminated cases
Population
Homoleptic and heteroleptic octahedral transition-metal complexes
Outcome
Predicted spin-state energetics and metal–ligand bond lengths

Structured fields used in claim comparison tables when every cited study has a complete layer.

What they did

They DFT-optimized homoleptic and heteroleptic octahedral complexes of Cr–Ni at 0–30% HF exchange, kept 2690 jobs after dropping spin-contaminated cases, and trained dropout ANNs on connectivity descriptors (connecting atom, ligand charge, denticity, metal/oxidation state). Test split was 40%; a CSD set checked transfer.

What they found

Spin-state RMSE is 3.0 (train) and 3.1 kcal mol–1 (test); ground state is right in 98% of test cases (528/538). Bond-length RMSE is 0.02 Å (LS) and 0.03 Å (HS test). On 35 CSD complexes MUE is 10 kcal mol–1 overall and 5 kcal mol–1 inside a Euclidean reliability cutoff; GGA-to-hybrid slope correction brings CSD MUE to 5 kcal mol–1.

The limits

What it doesn't show

The model is for octahedral first-row complexes with the ligand set used in training; cyclams and early metals are outliers (~30 kcal mol–1). Descriptors ignore 3D interligand clashes. Predictions target DFT (B3LYP-family) numbers, not experiment directly.

Key terms

Spin-state splitting ΔE(H–L)
Energy difference between high-spin and low-spin states of a TM complex.
Hartree–Fock exchange (aHF)
Fraction of exact exchange in a hybrid DFT functional; higher aHF favors high spin.
Dropout ANN
Neural net that randomly zeros nodes during training to reduce overfitting.
Graph descriptor
2D connectivity features (not Cartesian coordinates) describing the ligand around the metal.
CSD
Cambridge Structural Database of experimental crystal structures used as an out-of-sample test.

Flashcards

1 / 11

Research intelligence for this paper

See its role on concept claims, tensions it is part of, placement history, and related discoveries.

Open paper intelligence

Quiz yourself

1 / 7

The filtered DFT training corpus contains how many geometry optimizations?

Common questions

Why avoid 3D Coulomb-matrix inputs?

So you can screen complexes before a geometry exists; connectivity still transfers.

How accurate is spin splitting on the held-out DFT set?

RMSE ~3 kcal mol–1; 98% correct ground-state labels.

When should you not trust the ANN?

If the Euclidean distance to training space is large (e.g. cyclams); errors then reach tens of kcal mol–1.

What else can the net predict?

M–L bond lengths and HF-exchange sensitivity, used to shift GGA splittings toward hybrids.

More on Computational chemistry