Atmospheric drag remains one of the main sources of uncertainty in low Earth orbit propagation because the thermospheric density depends strongly on solar and geomagnetic activity and can vary substantially over time and space. This uncertainty directly affects orbit determination, orbit propagation and conjunction assessment and is therefore highly relevant for operational applications such as space traffic management. At the same time, the density model used in such applications must not only be sufficiently accurate, but also computationally efficient and compatible with near-real-time data availability constraints. This creates a practical trade-off between physical fidelity, operational robustness and computational cost.

This thesis investigates whether a neural network-based correction to the JB2008 empirical thermospheric density model can improve atmospheric density estimation and, more importantly, whether such improvements translate into better orbit propagation performance. Rather than replacing the empirical model entirely, a hybrid correction framework is developed in which JB2008 provides the baseline density estimate and a lightweight feed-forward neural network learns a correction using baseline density, solar, geomagnetic, spatial and temporally embedded input features. This design is motivated by
the desire to preserve the physical and operational utility of the empirical baseline while still allowing the model to capture systematic, time-varying deviations linked to changing space weather conditions.

To support operational relevance, the methodology is explicitly designed with near-real-time use in mind. The input feature set includes solar and geomagnetic drivers, physically motivated temporal embeddings that represent thermospheric memory effects and trust score features that quantify the reliability of outage-prone high-cadence space weather inputs. The model is trained on accelerometer-derived density data from the CHAMP, GOCE and GRACE missions. A key methodological choice is the use of a simple feed-forward architecture rather than a recurrent sequence model. By embedding
the relevant temporal history directly into the input features, the approach remains compatible with fragmented datasets and avoids dependence on continuous sequences, while still allowing delayed thermospheric responses to be represented.

Model performance is evaluated in two complementary stages. In stage A, the neural network is validated directly against accelerometer-derived density estimates. The results show that the correction model reduces the mean and spread of the log-residual error distribution relative to JB2008 and yields
reconstructed density time series that better match the reference densities in both phase and amplitude. These findings indicate that the model learns more than a static offset correction and captures part of the time-varying thermospheric variability that is not fully represented by the empirical baseline.
In stage B, the corrected density model is assessed in orbit propagation against Swarm A precise orbit determination data over seven validation arcs spanning quiet, moderate and severe geomagnetic conditions. The neural network correction model achieves the lowest overall 24-hour along-track RMS error among the tested density configurations and performs best during disturbed periods, when empirical density models tend to degrade most strongly. This suggests that the learned correction is not only beneficial when reconstructing density, but can also improve orbit propagation performance.

Overall, the results of this thesis indicate that a lightweight neural-network correction to JB2008 can provide a promising and operationally relevant improvement to empirical thermospheric density modelling. In particular, the findings suggest that a comparatively simple feed-forward architecture, when
combined with physically informed temporal feature design and explicit data-reliability indicators, is already capable of learning useful corrections to an empirical baseline. At the same time, the final orbit propagation validation is limited to a single satellite and a restricted set of validation arcs. The results
should therefore be interpreted as a promising proof of concept rather than as definitive evidence of broad operational superiority across all low Earth orbit regimes. Broader POD-based validation across additional satellites, altitude regimes and operational scenarios is required before more general conclusions can be drawn.

The use of plasma actuators for active flow control presents an interesting option for the design of an aircraft. Plasma-actuated systems boast the opportunities of having facilitated maintenance, increased operational efficiency, and a decreased noise profile, with minimal impact on general performance. Yet, due to the novelty of the technology, the exact implementation of a 100% plasma-controlled drone is yet to be done. Group 16 was tasked with designing an unmanned aerial vehicle (UAV) which omits conventional mechanically-actuated control surfaces, and instead uses plasma actuators. This report explores the various technical and non-technical design facets of the Plasma-actuated Unmanned Light Surveillance and Eco-friendly Drone (PULSE Drone). Due to the inherent novelty in the design stages, this paper aims to explain the current PULSE Drone design and its approach such that it becomes a benchmark for future development. From the depth of explanations, this paper becomes a relevant source for future projects aiming to incorporate plasma actuators.