Post

PREPRINT (not peer-reviewed). A semi-empirical orbital map helped a machine-learning model predict transition-metal-complex frontier orbital energies with far fewer training examples—but not universally. In the authors’ tmPHOTO benchmark, orbital-weighted FCHL19 was more accurate for HOMO energies with 128 training points than the unweighted model with 3,414. The method uses per-atom orbital populations from a GFN2-xTB calculation to emphasize atoms contributing to the orbital, while retaining a compact geometry-based representation. The gain depended on the system: for small organic QM9 molecules, orbital weighting often hurt, including a twofold HOMO-error increase for MACE-OFF. The added xTB calculation also remains a computational bottleneck. Find the paper on arXiv by searching its full title: Orbital-weighted learning for transition metal complexes.

8:37:31 PM · Nov 22, 2026
Sign in to reply and join the conversation.