Crack detection, length estimation, and Remaining Useful Life (RUL) prediction are among the most studied topics in reliability engineering. Several research efforts have studied physics of failure (PoF) of different materials, along with data-driven approaches as an alternative to the traditional PoF studies. To bridge the gap between these two techniques, we propose a novel hybrid framework for fatigue crack length estimation and prediction. Physics-based modeling is performed on the fracture mechanics degradation data by estimating parameters of the Paris Law, including the associated uncertainties. Crack length estimations are inferred by feeding manually extracted features from ultrasonic signals to a Neural Network (NN). The crack length prediction is then performed using the Particle Filter (PF) approach, which takes the Paris Law as a move function and uses the NN's output as observation to update the crack growth path. This hybrid framework combines machine learning, physics-based modeling, and Bayesian updating with promising results.