AI Based Velocity Calibration for Cassini's Cosmic Dust Analyzer (CDA)

Abstract This paper presents a novel AI‐based velocity calibration method for the Cosmic Dust Analyzer (CDA) onboard the Cassini spacecraft, in particular for impacts on the large Impact Ionization Detector (IID). The legacy polynomial calibration function, derived from laboratory measurements at a dust accelerator facility, assigns a velocity‐independent relative uncertainty of and is restricted to signals with a peak amplitude above for the QI channel. We propose a two‐stage deep learning architecture that combines a Temporal Convolutional Network Autoencoder for unsupervised feature extraction with a Mixture Density Network for probabilistic velocity prediction. The Autoencoder compresses the 510‐sample QI impact signals into a 3‐dimensional latent space that captures the essential morphological characteristics of each impact event, while the Mixture Density Network maps this latent representation to a conditional Gaussian distribution, providing both a mean velocity estimate and an individual, signal‐dependent uncertainty. Trained and validated on 839 laboratory calibration events from two measurement campaigns, the method achieves a mean relative uncertainty of approximately for velocities above , a substantial improvement over the legacy function. Furthermore, the amplitude threshold for evaluating a signal is reduced from 20 fC to 9.7 fC, significantly expanding the population of calibratable dust impacts in the Cassini CDA archive. The Autoencoder's reconstruction loss additionally serves as an anomaly detector, flagging signals that fall outside the learned calibration space. These improvements enable more accurate dust velocity measurements with direct implications for reanalyzing dust dynamics and orbital elements throughout the Saturnian system and beyond.