Machine-learning-based identification for initial clustering structure in relativistic heavy-ion collisions

$\ensuremath{\alpha}$-clustering structure is a significant topic in light nuclei. A Bayesian convolutional neural network (BCNN) applied to classify initial nonclustered and clustered configurations, namely Woods-Saxon distribution three-$\ensuremath{\alpha}$ triangular (four-$\ensuremath{\alpha}$ tetrahedral) for $^{12}\mathrm{C}$ $(^{16}\mathrm{O})$, from heavy-ion collision events generated within multiphase transport (AMPT) model. Azimuthal angle transverse momentum distributions of charged pions are taken as inputs train the classifier. On multiple-event basis, overall classification accuracy can reach $95%$ $^{12}\mathrm{C}/^{16}\mathrm{O}+^{197}\mathrm{Au}$ at $\sqrt{{S}_{NN}}=200\phantom{\rule{4pt}{0ex}}\mathrm{GeV}$. With proper constructions samples, predicted deviations on mixed samples with different proportions both configurations could be $5%$. In addition, setting simple confidence threshold further improve predictions dataset. Our results indicate promising extensive possibilities application machine-learning-based techniques real data some other problems physics collisions.