Atmospheric Modeling, Data Assimilation and Predictability

Introduction to atmospheric predictability

In his 1951 paper on NWP, Charney indicated that he expected that even as models improved there would still be a limited range to skillful atmospheric predictions, but he attributed this to inevitable model deficiencies and finite errors in the initial conditions. Lorenz (1963a, b) discovered the fact that the atmosphere, like any dynamical system with instabilities, has a finite limit of predictability (which he estimated to be about two weeks) even if the model is perfect, and even if the initial conditions are known almost perfectly. He did so by performing what is now denoted an “identical twin” experiment: he compared two runs made with the same model but with initial conditions that differed only very slightly. Just from round-off errors, he found that after a few weeks the two solutions were as different from each other as two random trajectories of the model.

Lorenz (1993) described how this fundamental discovery took place: His original goal had been to show that statistical prediction could not match the accuracy attainable with a nonlinear dynamical model, and therefore that NWP had a potential for predictive skill beyond that attainable purely through statistical methods. He had acquired a Royal-McBee LGP-30 computer, with a memory of 4K words and a speed of 60 multiplications per second, which for the late 1950s was very powerful.