Radiation-induced photocurrent effects represent a threat to microelectronic components operating in space, manmade and terrestrial radiation environments. Analysis of these threats by circuit simulations requires accurate and computationally efficient compact models. Most existing compact models are based on closed form analytic solutions of the governing equations and require empirical assumptions and idealizations that can limit their validity. In this paper we formulate an alternative numerical, data-driven approach that learns a compact model from data representative of the type of measurements one can obtain in an experimental facility. To develop the model we start from a generic discrete-time dynamical system and then use physics knowledge to refine its structure. Numerical studies demonstrate the potential of the model and establish some empirical guidelines for its training.