Systems and methods for machine-learning-based aircraft icing prediction use supervised and unsupervised learning to process real-time environmental data, such as onboard measurements of outside air temperature and dew point, to predict a risk of icing and determine whether to issue an icing risk alert to an onboard crewmember or a remote operator, and/or to recommend an icing avoidance maneuver. The systems and methods can use reinforcement learning to generate a confidence metric in the predicted risk of icing, to determine a time or distance to predicting icing, and/or to not issue an alert or recommend a maneuver in consideration of historical data in a “library of learning” and/or other flight data such as airspeed, altitude, time of year, and weather conditions. The predictive systems and methods are low-cost and low-power, do not require onboard weather radar, and can be effective for use in smaller aircraft that are completely icing-intolerant.