Transfer Learning for Predictive Molecular Simulations: Data-Efficient Potential Energy Surfaces at CCSD(T) Accuracy

Accurate potential energy surfaces (PESs) are critical for predictive molecular simulations. However, obtaining a PES at the highest levels of quantum chemical accuracy, such as CCSD(T), becomes computationally infeasible as molecular size increases. This work presents CCSD(T)-quality PESs using data-efficient techniques based on transfer learning to obtain state-of-the-art accuracy at a fraction of the computational cost for systems that would otherwise be intractable. Most importantly, the framework for accurate molecular simulations pursued here extends beyond specific observables and follows a rational strategy to obtain highest-accuracy PESs, which can be used for applications to spectroscopy and other experiments. As rigorous tests of the PESs, semiclassical tunnelling splittings for tropolone and the (propiolic acid)-(formic acid) dimer (PFD) as well as anharmonic frequencies for tropolone were determined. For tropolone, all observables are in excellent agreement with the experiment using the high-level PES, whereas for PFD, the agreement is less good but still orders of magnitude better than previous calculations.