Flashcard library · Chemistry · PaperFren
Neural nets predict TM spin states and bonds
Graph-based ANNs trained on 2690 DFT geometries of Cr–Ni octahedral complexes predict spin splitting and metal–ligand distances without 3D coordinates.
Keep what you just learned
Save these 11 cards to your NoteFren library. Your first review starts right away, and each card links back to the study it came from.
What metals are in the training complexes?
First-row octahedral Cr, Mn, Fe, Co, and Ni in common oxidation states.
Source: Predicting electronic structure properties of transition metal complexes with neural networks · DOI · quote verified
How many DFT geometries survived filtering?
2690 of a theoretical 2856.
Source: Predicting electronic structure properties of transition metal complexes with neural networks · DOI · quote verified
Are the ANN inputs 3D coordinates?
No; truncated graph/connectivity descriptors of the ligands.
Source: Predicting electronic structure properties of transition metal complexes with neural networks · DOI · quote verified
Train/test split used?
60% train, 40% test, order randomized.
Source: Predicting electronic structure properties of transition metal complexes with neural networks · DOI · quote verified
Spin-splitting RMSE on train vs test?
3.0 and 3.1 kcal mol–1.
Source: Predicting electronic structure properties of transition metal complexes with neural networks · DOI · quote verified
Ground-state classification accuracy on the test set?
98% (528 of 538); misses are near-degenerate DFT cases.
Source: Predicting electronic structure properties of transition metal complexes with neural networks · DOI · quote verified
Bond-length test RMSE?
0.02 Å (low-spin) and 0.03 Å (high-spin).
Source: Predicting electronic structure properties of transition metal complexes with neural networks · DOI · quote verified
How does HF exchange affect spin?
Higher exact exchange generally favors high-spin; the net also predicts that slope.
Source: Predicting electronic structure properties of transition metal complexes with neural networks · DOI · quote verified
CSD transfer error without filters?
MUE 10 kcal mol–1 and RMSE 13 kcal mol–1 on 35 structures.
Source: Predicting electronic structure properties of transition metal complexes with neural networks · DOI · quote verified
Which CSD ligands fail badly?
Early-metal cyclams, overestimated by about 30 kcal mol–1.
Source: Predicting electronic structure properties of transition metal complexes with neural networks · DOI · quote verified
How can GGA results be shifted toward hybrids?
Multiply the exchange-sensitivity ANN slope by ΔaHF; CSD MUE falls to 5 kcal mol–1.
Source: Predicting electronic structure properties of transition metal complexes with neural networks · DOI · quote verified