Speaker
Description
Modeling the Martian nightside thermosphere is challenging due to the absence of direct solar illumination and highly irregular in-situ sampling. We present a multi task physics informed neural network (MT-PINN) trained on more than a decade of MAVEN/NGIMS observations (MY 32-38) to reconstruct the densities of O, CO₂, N₂, and Ar. To represent coupling with space weather drivers, the model incorporates solar wind parameters with hour time lags, implicitly accounting for possible delays in dayside-to-nightside energy transfer.
Unlike purely data driven regressors that often produce non physical density inversions in sparsely sampled regions, the architecture introduces a weak monotonicity prior via automatic differentiation. This constraint penalizes positive vertical gradients in logarithmic density, encouraging behavior consistent with hydrostatic expectations while preserving the model’s ability to capture small scale variability.
Validation using a strict orbit disjoint split shows that the physics informed regularization substantially reduces inversion rates (up to 90% for certain species) while maintaining high predictive accuracy (RMSE ≈ 0.22). The resulting differentiable surrogate enables rapid profile reconstruction along arbitrary trajectories and provides an efficient tool for exploring the sensitivity of the Martian upper atmosphere to solar and heliospheric forcing.