Coupled cluster finite temperature simulations of periodic materials via machine learning

Density functional theory (DFT) is widely used for materials simulations, but results can vary with the exchange–correlation functional, limiting predictive power. Coupled cluster theory with singles, doubles, and perturbative triples [CCSD(T)] is considered a gold standard for non-strongly correlated systems, yet its cost hinders applications to periodic materials and especially finite-temperature properties that require molecular dynamics (MD). Here, we combine an efficient periodic CCSD(T) implementation with data-efficient machine learning (ML), thermodynamic perturbation theory (TPT), and Monte Carlo (MC) sampling to enable finite-temperature simulations at post-Hartree–Fock accuracy using only a limited number of high-level single-point calculations. We demonstrate the approach by computing the enthalpy of adsorption of CO2 in protonated chabazite (a zeolite), showing that ML-accelerated CCSD(T) yields results in excellent agreement with experiment while drastically reducing computational cost. Our proof-of-principle paves the way for broader use of highly accurate correlated methods in materials simulations at finite temperature.