Transferability of machine-learned functionals beyond their training chemistry
In plain words
Functionals fitted by machine learning to accurate data now beat hand-made ones on molecules similar to their training set. It is unknown whether they stay accurate for metals, transition-metal compounds and solids they never saw.
Precise statement
For a machine-learned exchange-correlation functional trained only on main-group molecular data, determine its errors on out-of-distribution sets: 3d transition-metal reaction energies and spin gaps, lattice constants and cohesive energies of metals, and band gaps of solids, relative to the best conventional hybrid. An answer is a quantitative transfer test, or a training principle (exact constraints, data types) shown to guarantee transfer.
What would settle it
Blind evaluation of trained functionals on held-out transition-metal and solid-state benchmarks with reference data from coupled cluster, quantum Monte Carlo and experiment.
Status in the literature
Unverified note
Learned functionals released from 2021 to 2025 improved main-group thermochemistry; transfer to transition metals and solids is not established.