DFTB is a high-speed train in the world of quantum chemistry: hundreds of times faster than classical DFT, yet still on electronic rails. But its engine — iterative self-consistency of charges — sometimes resembles an orchestra where each instrument starts from an arbitrary note. When initial charges are zero, the system takes a long time to tune up: dozens, maybe hundreds of cycles, until the dissonance of electrostatic interactions gives way to harmony. For transition metal oxides or interfaces, this symphony risks becoming eternal — many structures never converge to the energy minimum.

Machine learning comes to the rescue. Instead of a blind start from zero, the model predicts nearly converged charges using the molecule’s geometry. Each atom gets a ‘chemical passport’ — a SOAP descriptor that encodes the types of neighbors and distances to them, staying invariant under rotations and permutations. Kernel ridge regression, trained separately for each element, captures its characteristic signature in the distribution of electrostatic charge — much like an experienced chemist predicts the most electronegative atoms from structure. Only the machine does it without hesitation and with accuracy down to hundredths of a charge. This warm start is then fed into the standard SCC cycle. Here, the Hamiltonian matrix elements depend on charges: H = H⁰ + ½ S ∑ ($γ_{IK+}$$γ_{JK}$) Δq_K, and the total energy includes an electronic part and a Coulombic contribution: $E_{SCC}$ = ∑ $f_i$ ⟨$ψ_i$|Ĥ₀|$ψ_i$⟩ + ½ ∑ $γ_{IJ}$ Δq_I Δq_J + ∑ $V^rep_{IJ}$. So ML doesn’t replace physics; it just sets the right key before the symphony begins.

The test was conducted on six chemical universes: from light organics with hydrogen, carbon, and oxygen to water clusters, biomolecules, nickel oxides, and the solid electrolyte LLZO — a candidate for lithium batteries. The charge error is negligible — 0.001–0.01 e, but the effect is colossal: iterations decreased by 84% for nickel oxides (from 62.5 to 10.0), by 58% for LLZO, by 34% for water clusters. But the most dramatic result is resurrection: 891 nickel oxide structures out of 1118, previously stuck in an endless numerical loop, finally achieved convergence.

This work is not a victory of black boxes over physics, but an elegant symbiosis. The hybrid scheme ‘ML + quantum algorithm’ removes a headache from high-throughput screening: manual tweaking of initial conditions used to choke automated pipelines, but now a warm start gives the green light. The next step is models that predict not only total charge but also the filling of individual shells (s, p, d) for transition metals. Universal ‘charge oracles’, trained on giant chemical databases, promise to become an everlasting starter for any parameters; neural networks might even capture long-range effects. Thus we approach a world where the design of catalysts, drugs, and batteries becomes fast and reliable engineering routine.