|
Résumé |
We investigate the advantages of developing machine learning methods
in the context of theoretical nuclear physics and in particular of nuclear
fission.
Actually, the description of the fission realized with the HFB+TDGCM
is numerically very expensive. Because the method requires a potential
energy surface (PES), for each nucleus, made of thousands of HFB states
at different nuclear deformations. These latters are obtained by solving
the HFB equation under fixed deformation constraints and by varying
the oscillator basis parameters with respect to the minimum of energy.
In contrast to the standard methods, we propose to optimize the basis
parameters thanks to global optimization algorithms with the HFB en-
ergy modelized by a Gaussian process. This Bayesian approach offers the
estimation of the energy, its associated error and it takes into consider-
ation the numerical noise. We observe ameliorations for a single HFB
calculation, by reaching a lower energy faster than classical minimization
methods. Furthermore for multiple HFB calculations, such as the PES
production, we are able to improve the distribution of the calculations to
be performed simultaneously. Quantitatively, one notices a speedup of 5
times for the production of the PESs in contrast to the reference.
Finally, we will present the learning of the nuclear properties (defro-
mations, energy, etc...) from calculated PESs, in order to construct an
approximated HFB state as a starting point of a complete HFB calcula-
tion. This method is very helpful in spite of improving the production of
new PESs by compiling all the knowledge obtained from calculated nuclei.
Also, this is a big step in solving scalability problems, so far limited by a
propagation mechanism. |
