Models for the evolution of hidden microstructural states are needed for fast prediction and closed-loop control of workpiece properties. Machine learning allows to obtain models by learning from experimental data, avoiding the limitations of explicitly defined physics-based models. However, the identification of the parameters of deep network structures, reliable extrapolation and fast online assimilation to new measurements are open problems. At the example of titanium forging, a new approach is investigated that combines a hybrid physics-informed microstructure and flow stress model that draws upon a long short-term memory network with a particle filter for online data assimilation to new measurements.