Skip to content
PaperFren

Machine learning · Electronic structure

A neural network predicted spin states well — and reported where it should not be trusted

Evidence: EmergingMore than one study points the same way, but the body is still thin. What the labels mean

Study published Jan 1, 2017. PaperFren added this explanation Sep 20, 2026.

Save this development or follow its topic to track what changes.

Short answer

Connectivity-trained neural networks predicted transition-metal spin states to about 3 kcal mol⁻¹ within their domain, with a distance-based cutoff separating reliable predictions from 30 kcal mol⁻¹ failures.

What happened

Janet and Kulik optimised homoleptic and heteroleptic octahedral complexes of chromium through nickel at 0–30% Hartree-Fock exchange, kept 2,690 calculations after removing spin-contaminated cases, and trained dropout artificial neural networks on connectivity descriptors. Spin-state splitting root-mean-square error was 3.0 kcal mol⁻¹ in training and 3.1 in test, with the correct ground state in 528 of 538 test cases. Bond-length errors were 0.02–0.03 Å. On 35 complexes from the Cambridge Structural Database the mean unsigned error was 10 kcal mol⁻¹ overall and 5 within a Euclidean reliability cutoff, while cyclams and early metals fell around 30 kcal mol⁻¹ out.

Why it matters

A model's headline accuracy is measured inside its training distribution, which is where it will not be used. The reliability cutoff is the transferable contribution here: it converts 'the model is accurate' into 'the model is accurate for inputs like these', which is the form a working chemist can act on.

Evidence

Study type
Supervised learning on DFT-generated data with held-out test and external structural-database evaluation
Sample
2,690 retained calculations; 538 test cases; 35 external complexes
Journal
Chemical Science · peer reviewed
Replication
Not assessed in this corpus
Limitations
Restricted to octahedral first-row complexes with the trained ligand set. Connectivity descriptors ignore three-dimensional clashes, and predictions inherit whatever errors the underlying DFT carries.

What this connects to

Sources

The one study this explanation is built from, by the role each plays. Every source links to PaperFren’s explanation of it and to the original paper.

Primary study

  • Neural nets predict TM spin states and bonds

    Janet JP, Kulik HJ · 2017 · Chemical science · 114 citations

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

    What it does not 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.

    PaperFren explanationStudy with cards and a quizOriginal paper (DOI)cc by

Before

Spin-state energetics for transition-metal complexes required electronic structure calculations per complex, and machine-learning surrogates were typically reported with a single accuracy figure.

Now

A surrogate reaches useful accuracy and ships with a domain-of-applicability test. The model covers octahedral first-row complexes with the trained ligand set, descriptors ignore three-dimensional interligand clashes, and cyclams and early metals are explicit outliers.