Conceptual framework for the application of deep neural networks to surface composition reconstruction from Mercury's exospheric data

TL;DR

This work proposes a conceptual framework to infer Mercury’s surface composition from exospheric measurements by training a supervised deep neural network (an MLP) on simulated data that encapsulates the surface, exosphere, and environmental drivers. The model learns to estimate surface elemental fractions from exospheric densities, with a data-generation model incorporating micrometeorite vaporization, ion sputtering, thermal desorption, and photon-stimulated desorption, plus planetary geometry and solar conditions. Training and testing show the MLP can reconstruct surface maps with high fidelity at low altitudes (e.g., ES4 ≈ 89.7%, R^2 ≈ 83.4% at 200 km), especially under daylight conditions, demonstrating the framework’s viability as a data-driven estimator of exosphere-generation mechanisms. Although the study uses synthetic data and a simplified magnetic field, it lays the groundwork for applying the method to real SERENA data from BepiColombo and for extending to more realistic process models and temporal dynamics.

Abstract

Surface information derived from exospheric measurements at planetary bodies complements surface mapping provided by dedicated imagers, offering critical insights into surface release processes, interactions within the planetary environment, space weathering, and planetary evolution. This study explores the feasibility of deriving Mercury's regolith elemental composition from in-situ measurements of its neutral exosphere using deep neural networks (DNNs). We present a supervised feed-forward DNN architecture - a multilayer perceptron (MLP) - that, starting from exospheric densities and proton precipitation fluxes, predicts the chemical elements of the surface regolith below. It serves as an estimator for the surface-exosphere interaction and the processes leading to exosphere formation. Because the DNN requires a comprehensive exospheric dataset not available from previous missions, this study uses simulated exosphere components and simulated drivers. Extensive training and testing campaigns demonstrate the MLP's ability to accurately predict and reconstruct surface composition maps from these simulated measurements. Although this initial version does not aim to reproduce Mercury's actual surface composition, it provides a proof of concept, showcasing the algorithm's robustness and capacity for handling complex datasets to create estimators for exospheric generation models. Moreover, our tests reveal substantial potential for further development, suggesting that this method could significantly enhance the analysis of complex surface-exosphere interactions and complement planetary exosphere models. This work anticipates applying the approach to data from the BepiColombo mission, specifically the SERENA package, whose nominal phase begins in 2027.

Figures (21)

• Figure 1: DNN prediction task schematics. The input to the neural network on the left are the exospheric densities at a single location in the exosphere. The output is the relative surface elemental composition as fractions summing up to 1 at a surface area just below the exospheric measurement. The hidden layer box consists of multiple layers and represents the complex, often nonlinear, relationships between the inputs and the outputs of the neural net.
• Figure 2: MLP DNN architecture overview. The data generation model produces both the inputs and the outputs for training, validating, and testing the algorithm. This simulated data is passed through the MLP DNN model in the training, hyperparameter tuning and testing phases, respectively. The backpropagation optimization uses the loss function, regularizer and optimization algorithm to adjust the weights (internal parameters) of the neural network. In a separate process, the hyperparameter tuner adjusts/optimizes the MLP DNN by minimizing the errors on the validation dataset. After the final training, the previously unseen data from the testing sets is passed through the network and the accuracy of the predictions (performance of the network) is evaluated. Note: This proof-of-concept study trains and tests the DNN's capability to estimate the processes that are simulated via the Data Generation Model. As one of the goals of future studies will be to construct a data-driven representation of real physical processes, one approach in that case would be to replace the Data Generation Model with the real physical data generation mechanisms. Other alternative approaches are discussed in Section 6.
• Figure 3: MLP DNN basic structure schematics. Exospheric densities form the input layer on the left. The output layer is formed from the relative surface elemental composition as fractions summing up to 1. There are $N$ number of hidden layers with $n_1$ to $n_N$ number of neural units (neurons). This structure represents the relationships between the inputs and the outputs of the MLP. The neural synapses, connections between the neural units, form the weight (DNN parameter) matrices $W_1$ to $W_{N+1}$.
• Figure 4: Maps of incidence angles due to planet orientation at two consecutive perihelia. Perihelion 1 on the left and perihelion 2 on the right. The top panels represent the incidence maps of solar photons on the surface. The middle panels represent the angles between the planet's velocity and the surface normal. The bottom panels show the magnetic field cusp footprints on the surface. A shift of 180 degrees in longitude between the two perihelia is notable due to the spin-orbit resonance of Mercury.
• Figure 5: Learning curve for the MLP DNN training. The blue and orange curves show the evolution of the average prediction similarity of the full training dataset (300 subsets, 1,360,800 data points) and the hold-out validation dataset (1 subset, 648 data points) respectively.