1: \begin{abstract} 

2: Machine Learning (ML) has become a promising tool for improving the

3: quality of atomistic simulations. Using formaldehyde as a benchmark

4: system for intramolecular interactions, a comparative assessment of ML

5: models based on state-of-the-art variants of deep neural networks

6: (NN), reproducing kernel Hilbert space (RKHS+F), and kernel ridge

7: regression (KRR) is presented. Learning curves for energies and atomic

8: forces indicate rapid convergence towards excellent predictions for

9: B3LYP, MP2, and CCSD(T)-F12 reference results for modestly sized (in

10: the hundreds) training sets. Typically, learning curve off-sets decay

11: as one goes from NN (PhysNet) to RKHS+F to KRR (FCHL). Conversely, the

12: predictive power for extrapolation of energies towards new geometries

13: increases in the same order with RKHS+F and FCHL performing almost

14: equally. For harmonic vibrational frequencies, the picture is less

15: clear, with PhysNet and FCHL yielding respectively flat learning at

16: $\sim 1$ and $\sim 0.2$ cm$^{-1}$ no matter which reference method,

17: while RKHS+F models level off for B3LYP, and exhibit continued

18: improvements for MP2 and CCSD(T)-F12. Finite-temperature molecular

19: dynamics (MD) simulations with the same initial conditions yield

20: indistinguishable infrared spectra with good performance compared with

21: experiment except for the high-frequency modes involving hydrogen

22: stretch motion which is a known limitation of MD for vibrational

23: spectroscopy. For sufficiently large training set sizes all three

24: models can detect insufficient convergence (``noise'') of the

25: reference electronic structure calculations in that the learning

26: curves level off. Transfer learning (TL) from B3LYP to CCSD(T)-F12

27: with PhysNet indicates that additional improvements in data efficiency

28: can be achieved.

29: \end{abstract}

30: