“…While still recent overall, the evidence for the applicability of ML models to atomistic simulation and electronic property predictions has become overwhelming by now.A n incomplete list of studies which infer QM observables,rather than solving the conventional quantum chemistry approximations to Schrçdingerse quation (SE), [29] includes ML models of enthalpies of formation, [30,31] density functionals, [32] basis-set effects, [33] reorganization energies, [34] chemical reactivity, [35] atomization energies, [11] kinetic density functionals, [36] electronic ground-state properties, [20,37] transition state dividing surfaces, [38] polymer properties, [39] electron transmission coefficients, [40] crystal properties, [41][42][43][44] Anderson impurity models, [45,46] NMR nuclear shifts, [47] frontier orbital eigenvalues, [13] atomic charges,d ipole and quadrupole-moments, [48] electronic excitation energies, [49] Møller-Plesset theory corrections, [50] interatomic many-body expansions, [51] transition metals, [52] bonds in molecules, [53] and electron density functionals for use in ab initio molecular dynamics. [54] But also fitting the potential energy surface is awell-known problem in quantum chemistry,and has triggered substantial research on how to interpolate best with the least number of single point calculations.W agner, Schatz and Bowman introduced the modern computing perspective in the 80s, [55,56] while neural network fits to PES started to appear with Sumpter and Noid in 1992, [57] followed by Lorenz, Gross and Scheffler in 2004, [58] Manzhos and Carrington in 2006, [59] Behler and Parrinello in 2007,…”