Machine Learning Electron Correlation Via the 2-Electron Reduced Density Matrix

M. Pavanello

Machine learning models for correlated electronic states are extremely powerful. In principle, they capture the full complexity of many-body physics and are therefore universally applicable. In practice, they dramatically reduce the computational cost of obtaining approximate yet accurate correlated electronic structures for benchmark simulations. We demonstrate that targeting the two-body reduced density matrix (2-RDM) yields size-extensive and tightly converged ML models. Their applicability spans a wide range of domains: from condensed-phase chemistry and reaction barrier calculations to high-accuracy X-ray scattering experiments.