Quantum-Aided Gravity Gradient Estimation

Built a deep learning model to estimate vertical gravity gradients from anomaly data, overcoming computational bottlenecks in quantum-aided map-matching navigation.

Problem

Gravity map-matching using cold-atom quantum sensors is highly jam-resistant but computationally expensive. Evaluating gravity gradients at candidate particle locations in a particle filter traditionally requires calculating heavy surface integrals, presenting a significant bottleneck for real-time applications.

Approach

• Surrogate Modeling: Developed a 6-layer fully connected deep neural network (DNN) using PyTorch to approximate the surface integral.
• Hyperparameter Optimization: Utilized the Optuna library to execute a 250-trial optimization sweep, refining hidden layers, learning rates, dropouts, and learning rate decay schedulers (CosineAnnealingLR).
• Real-Time Integration: Integrated the DNN surrogate model directly into a simulated particle filter framework, allowing vertical gravity gradients to be predicted in a single forward pass.

Key Highlights

• Computation Speedup: Reduced per-sample inference time by approximately 14% with significantly lower variance (jitter) compared to numerical integration.
• High Precision: Maintained sub-nanogal (under $8.02 \times 10^-9$ s$^-2$) root-mean-square error, well below the noise floor of current quantum sensors.
• Robust Navigation: Tested in simulated 22.5 km flight trajectories, keeping positioning errors bounded within 35.5 meters.