QiankunNet: A Transformer‑Based Framework for Solving the Many‑Electron Schrödinger Equation

Problem

The many‑electron Schrödinger equation determines all material properties, but its computational cost grows exponentially with the number of electrons, creating an “exponential wall” that limits traditional quantum‑chemical methods.

Method

A decoder‑only Transformer architecture was built to construct quantum wave functions. The model receives a sequence of electron‑orbital occupation numbers and outputs the wave‑function amplitude and phase. Because the whole pipeline is end‑to‑end differentiable, variational energy can be minimized directly by back‑propagation using the Variational Monte Carlo (VMC) method.

An autoregressive sampling scheme based on Monte‑Carlo Tree Search (MCTS) generates independent electron configurations in parallel, avoiding the high sample correlation and slow convergence of conventional samplers. A physics‑inspired initialization derived from configuration interaction (CI) further accelerates variational optimization.

Benchmarks

On molecular benchmarks containing up to 30 spin orbitals, the Transformer‑based QiankunNet achieves 99.9 % of the full‑configuration interaction (FCI) correlation energy. Compared with the coupled‑cluster method CCSD(T), QiankunNet shows superior accuracy on strong‑correlation problems such as bond dissociation. Relative to earlier neural‑network quantum‑state approaches, QiankunNet runs roughly ten times faster on a 30‑orbital system while delivering higher precision.

Application to Transition‑Metal Chemistry

The framework was applied to the Fenton reaction, modeling the complete O‑O bond cleavage pathway in the complex [Fe(H₂O)₅(H₂O₂)]²⁺. The calculation accurately captures the electronic‑structure evolution from Fe(II) to Fe(III), demonstrating the method’s capability for complex transition‑metal systems.

Implication

These results show that attention mechanisms, originally developed for natural‑language processing, can faithfully represent quantum wave functions, opening a pathway for integrating large‑language‑model architectures into core quantum‑chemical research.

Reference: Solving the many‑electron Schrödinger Equation with a Transformer‑based framework, Nature Communications (2025). https://www.nature.com/articles/s41467-025-63219-2