We develop an approach for fully Bayesian learning and calibration of spatiotemporal dynamical mechanistic models based on noisy observations. Calibration is achieved by melding information from observed data with simulated computer experiments from the mechanistic system. The joint melding makes use of both Gaussian and non-Gaussian state-space methods as well as Gaussian process regression. Assuming the dynamical system is controlled by a finite collection of inputs, Gaussian process regression learns the effect of these parameters through a number of training runs, driving the stochastic innovations of the spatiotemporal state-space component. This enables efficient modeling of the dynamics over space and time. Through reduced-rank Gaussian processes and a conjugate model specification, our methodology is applicable to large-scale calibration and inverse problems. Our method is general, extensible, and capable of learning a wide range of dynamical systems with potential model misspecification. We demonstrate this flexibility through solving inverse problems arising in the analysis of ordinary and partial nonlinear differential equations and, in addition, to a black-box computer model generating spatiotemporal dynamics across a network.