NN-xTB: density functional accuracy at semi empirical speed with neural network extended tight binding

Accurate prediction of molecular structure, energetics, and reactivity requires quantum chemistry methods that remain too costly for large-scale and high-throughput modeling. Semi-empirical tight-binding approaches are much faster, but their largely element-fixed parameters under-resolve local chemical-environment effects, limiting accuracy. We introduce a neural-network extension of extended-tight binding that learns bounded, environment-dependent shifts to Hamiltonian parameters, increasing chemical adaptivity while preserving self-consistency, charge and spin treatment, and correct long-range behavior. We show that, across benchmarks spanning main-group thermochemistry, noncovalent interactions, molecular forces, vibrational spectra and supramolecular complexes, this approach achieves density functional theory accuracy at near tight-binding cost. On the GMTKN55, the weighted mean absolute error decreases from 25.0 to 3.78 kcal/mol relative to GFN2-xTB; molecular force errors are lowest on eight of ten rMD17 molecules and improve by 10-40% relative to leading machine-learned interatomic potentials; and vibrational frequency errors decrease from 200.6 to 12.7 cm^-1, with runtime increasing by less than 20%.