Neural network biased corrections: Cautionary study in background corrections for quenched jets
David Stewart, Joern Putschke · Physical Review C · 2025
Jets clustered from heavy ion collision measurements combine a dense background of particles with those actually resulting from a hard partonic scattering. The background contribution to jet transverse momentum (${p}_{\mathrm{T}}$) may be corrected by subtracting the collision average background; however, the background inhomogeneity limits the resolution of this correction. Many recent studies have embedded jets into heavy ion backgrounds and demonstrated a markedly improved background correction is achievable by using neural networks (NNs) trained with aspects of jet substructure which are used to map measured jet ${p}_{\mathrm{T}}$ to the embedded truth jet ${p}_{\mathrm{T}}$. However, jet quenching in heavy ion collisions modifies jet substructure and correspondingly biases the NNs' background corrections. This study investigates those biases by using simulations of jet quenching in central $\mathrm{Au}+\mathrm{Au}$ collisions at $\sqrt{{s}_{\mathrm{NN}}}=200\phantom{\rule{0.16em}{0ex}}\mathrm{GeV}$ with hydrodynamically modeled quark-gluon plasma (QGP) evolution. To demonstrate the magnitude of the effect of such biases in measurement, a leading jet nuclear modification factor (${R}_{\mathrm{AA}}$) is calculated and reported using the NN background correction on jets quenched utilizing a brick of QGP.