A method of generating balanced, fast-growing perturbations to a nonlinear model trajectory for a given finite perturbation amplitude. The technique consists of generating a control run of a nonlinear model for a short period of time (e.g., 6 h), then perturbing the atmospheric initial conditions and running the same model again for the same period of time (perturbed run). The difference between the two model runs at the final time is adjusted to the amplitude of the initial perturbation and is added to the new control initial condition. The method is then repeated. After a few days of iteration, the difference between the control and perturbed model runs represents a sample of fast-growing nonlinear perturbations. Multiple breeding cycles, started with different arbitrary initial perturbations, provide a broader sample of fast-growing perturbations. The technique can also be applied to a series of atmospheric analysis fields, where the control forecast always starts from the latest available analysis. See bred vectors.