Dynamic nonlinear models of metabolism offer a significant advantage as compared to constraint-based stoichiometric descriptions. However, progress in the development of large-scale nonlinear models has been hindered by both structural and quantitative uncertainties. In particular, the knowledge about kinetic rate laws and their parameters is till today still very limited when compared to the number of stoichiometric reactions known to be present in a large-scale metabolic model. In addition, strategies to systematically identify and implement large-scale dynamic models for metabolism are still lacking. In this contribution, we propose a novel methodology for development of dynamic nonlinear models for metabolism. Using the ORACLE (Optimization and Risk Analysis of Complex Living Entities) framework, we integrate thermodynamics and available omics and kinetic data into a large-scale stoichiometric model. The resulting set of log-linear kinetic models is used to compute kinetic parameters of the involved enzymatic reactions such as the maximal velocities and Michaelis constants. These kinetic parameters are in turn used to compute populations of stable, nonlinear, dynamic models sharing the same stable steady-state as the log-linear ones. The computed models offer unprecedented possibilities for system analysis, e.g. to study the responses of metabolism upon large perturbations; to investigate time course evolutions in and around the steady state; and to identify multiple steady-states and their basins of attraction. We illustrate the features of the generated models in the case of optimally grown E. coli, where our analysis of the estimated maximal reaction rates highlights the significance of network thermodynamics in constraining the variability of these quantities.