Fig. 6: Power spectra and potential energy surface exploration via minima hopping. | Nature Communications

Fig. 6: Power spectra and potential energy surface exploration via minima hopping.

From: A Euclidean transformer for fast and stable machine learned force fields

Fig. 6

a Power spectra for the buckyball catcher (upper panel) and the double walled nanotube (lower panel) computed as the Fourier transform of the velocity auto-correlation function. For the nanotube, the structure is shown from the side and from the front. b Power spectrum for DHA from an NVE at the zero point energy (ZPE) (light blue) and 500 K (dark blue), as well as the frequencies from harmonic approximation (red dashed lines). c Results of a minima search for DHA. We ran the simulation until 10k minima had been visited, which corresponds to 20 M molecular dynamics steps for the escape trials and to  ~10 M PES evaluations for the structure relaxations, afterwards. Next to it minima with the largest energy (top), the lowest energy (bottom) and an example minimum with an intermediate energy value (middle) are depicted. d Ramachandran density plots for the training conformations (upper, blue) and of the visited minima during minima hopping (lower, green) for two of the six backbone angles in Ac-Ala3-NHMe. Yellow dots correspond to the actually visited minima. Parts of the visited minima have not been in the training data, hinting towards the capability of the model to find minima beyond the training data. e Ac-Ala3-NHMe structure with backbone angles as inset. f Relative energies for four minima, which have been selected from the regions in ψ − ϕ space visited most frequently during minima hopping (A–D in (d)). SO3krates energies are compared to a DFT single point calculation and to the conformation obtained from a full DFT relaxation starting from the minima obtained from SO3krates. g Location in the Ramachandran plot of the minima obtained with SO3krates and the relaxed DFT minima.

Back to article page