Development of a Method to Detect High Ice Water Content Environments Using Machine Learning

This paper describes development of a method for discriminating high ice water content (HIWC) conditions that can disrupt jet-engine performance in commuter and large transport aircraft. Using input data from satellites, numerical weather prediction models, and ground-based radar, this effort employs machine learning to determine optimal combinations of available information using fuzzy logic. Airborne in situ measurements of ice water content (IWC) from a series of field experiments that sampled HIWC conditions serve as training data in the machine-learning process. The resulting method, known as the Algorithm for Prediction of HIWC Areas (ALPHA), estimates the likelihood of HIWC conditions over a three-dimensional domain. Performance statistics calculated from an independent subset of data reserved for verification indicate that the ALPHA has skill for detecting HIWC conditions, albeit with significant false alarm rates. Probability of detection (POD), probability of false detection (POFD), and false alarm ratio (FAR) are 86%, 29% (60% when IWC below 0.1 g m(-3) are omitted), and 51%, respectively, for one set of detection thresholds using in situ measurements. Corresponding receiver operating characteristic (ROC) curves give an area under the curve of 0.85 when considering all data and 0.69 for only points with IWC of at least 0.1 g m(-3). Monte Carlo simulations suggest that aircraft sampling biases resulted in a positive POD bias and the actual probability of detection is between 78.5% and 83.1% (95% confidence interval). Analysis of individual case studies shows that the ALPHA output product generally tracks variation in the measured IWC.