Conceptual framework for the application of deep neural networks to surface composition reconstruction from Mercury's exospheric data (v2.4, focused study, obsolete revision)

Superseded by v2.5 A complete study with realistic environment models, multiple processes, single optimized DNN, focused on a single prediction task (elemental ratios at surface). This finalized paper is published as the Conceptual Framework of a second, more-focused study in the chronological collection of manuscripts exploring the conceptual application of machine-learning algorithms for the reconstruction of Mercury's surface composition from simulated exospheric measurements. Published also in arXiv. Single DNN type: An optimized MLP with 4 hidden layers. Elemental composition prediction (11 elements) from exospheric elemental measurements; maps reconstruction and analysis; metrics evaluations. Simulated dataset - realistic assumptions for the release processes (4 main processes - MIV, SP, PSD, TD), exospheric dynamics included; chemistry, magnetic field influences, and testing on actual Mercury surface and exosphere are not included.