Machine learning advent and derivative discontinuity of DFT functionals over gap state predictions among ACeO3 (A = Ba2+, Sr2+, Ca2+, Mg2+) proton conductors

[Display omitted] •Inconsistency in the gap-state evaluations using different DFT functionals due to lack of self-electron interaction correction within the Exc potential.•Ce4+ → Ce3+ facile reductions and hybridization differences alter the dual (valance and conduction) positions and the resultant electronic bandgap.•Material dependency and sensitivity of semiempirical Hybrid (HSE06) and DFT + U approach constitute erroneous band gap with exhaustive computational resource.•Machine learning approach with optimized dynamical batch assists accelerated convergence devoid of training instability and generalization loss. The electrophysical fluctuations within strongly correlated d and f-electron system such as ACeO3 (A = Ba2+, Sr2+, Ca2+, Mg2+) heavily relies upon the nature of chemical bonding, charge density distribution, dual-band positioning and the nature of hybridizations between the compositional constituents. Meanwhile, Ce4+ → Ce3+ facile reductions due to Ce-4f0 → Ce-4f1 electron occupancy additionally imparts band energy shifts with varying bandgaps. Besides implicit material characteristics, under and overestimated outcomes via distinct DFT functionals emerge due to inconsistent self-electron interactions within the EXC approximations. While fractional non-local Fock exchange within Hybrid (HSE06) reduce the artificial barrier to localization, material-dependent response of the functional and the sensitivity of Ueff within the DFT + U approach due to d-orbital positioning may invite erroneous bandgap estimations. Sophisticated and advanced double hybrid functionals (B2PLY, ωB97X-D) at higher computational expense also limits their practical utility to complex oxides. In this study, we illustrate the performance accuracy of the designed artificial neural network (ANN) towards the band energy predictions of distinct Ce-based proton conductors via optimized hyperparameters. The study also reflects upon the training instability and generalization loss as a function of dynamical batch size and the stochastic behaviour of the network corresponding to distinct input statistics.