Dear Abhishek,

The tools used to forecast weather, project likely climate scenarios, or to reanalyze data are all based on or derived from a General Circulation Model (GCM). These models require initial (IC) and boundary conditions (BC) before they can start to run forward and generate estimates of future situations.

Because of intrinsic difficulties (including inevitable uncertainties in the observational data, inconsistencies between those, and probably limitations in the model formulation too), the initial results are unreliable as the model attempts to stabilize. This is the 'spin up' period. Once the fields have adjusted and results become more stable, the model can be stimulated (forced) in a particular way and the results can be trusted, within limits that can themselves be documented.

Hence the spin up time may vary from model to model, and especially with the quality of the IC and BC, as well as with the purpose of the run. Given the cost of high performance computing, there is a trade-off between a long spin up period (to allow the model to 'forget' the IC and BC) and the need to quickly generate useful outcomes. The communities mentioned above (weather, climate, reanalysis) have generated ample materials on this matter.

Cheers, Michel.

Dear Abhishek,

The tools used to forecast weather, project likely climate scenarios, or to reanalyze data are all based on or derived from a General Circulation Model (GCM). These models require initial (IC) and boundary conditions (BC) before they can start to run forward and generate estimates of future situations.

Because of intrinsic difficulties (including inevitable uncertainties in the observational data, inconsistencies between those, and probably limitations in the model formulation too), the initial results are unreliable as the model attempts to stabilize. This is the 'spin up' period. Once the fields have adjusted and results become more stable, the model can be stimulated (forced) in a particular way and the results can be trusted, within limits that can themselves be documented.

Hence the spin up time may vary from model to model, and especially with the quality of the IC and BC, as well as with the purpose of the run. Given the cost of high performance computing, there is a trade-off between a long spin up period (to allow the model to 'forget' the IC and BC) and the need to quickly generate useful outcomes. The communities mentioned above (weather, climate, reanalysis) have generated ample materials on this matter.

Cheers, Michel.

Spin up time is simply the time taken for the computer model to approach its own climatology after being started from other initial conditions.  If a model were perfectly accurate and the initial conditions from which it is started were also perfect, then there would be no spin up time.  However, in practice, computer models of the atmosphere and ocean are imperfect and will drift from a given initial state towards their own preferred state.

For atmospheric models this is relatively fast as the atmosphere has a short memory (in most regions at least) and it normally takes about a month to spin up.  The only exception is the tropical lower stratosphere where the timescale is about a year.

Ocean models have a longer 'memory' and therefore take longer to spin up.  The near surface mixed layer will spin up within a year or so but the deep ocean takes many decades or even centuries.

By initialising models close to a realistic state the spin up time can be reduced but this requires a good set of initial conditions.

Best

Adam

To run a model, the initial conditions are required, either from observations or from other model outputs, or a combination of these two. The initial conditions need to be thermal-dynamically balanced associated with the model to be run. If not, the model will do an adjustment until the balance is reached. This adjustment is called model spinup, and the time needed is spinup time. Depends on the models, spinup time can be from only a few hours for a meso-scale weather numerical model, up to couple of years for a land surface model regarding soil moisture (Yang, Yang, M. Uddstrom, and M. Duncan (2011), Effects of short spin-up periods on soil moisture simulation and the causes over New Zealand, J. Geophys. Res., 116, D24108, doi:10.1029/2011JD016121).

I think all the above suggestions are useful. However, I would like to add an important tip if you plan to run several sensitivity experiments. Usually, every model has the capability to re-run the model, so for the first run give a maximum spin-up time (if you go long run also, that won't hurt the model and you at all), and plan to re-run the model for every sensitivity experiments starting from where you thought the model has enough spin-up. In this way, you can save a lot of computer resource and your valuable time !

Spin-up time for example, for an ocean model is the time taken for an ocean model to reach a state of statistical equilibrium under the applied forcing. But it is usually difficult for basin-scale and global general circulation models to reach this state, as it can take hundreds of years.

Give all the nice explanations, I would mention one example which i understood

If you choose the spin up time twice the total run time, the model will loop through all steps twice and it will use the the final concentrations from the last step of the first run to to initialize the first step of the second run. This is useful for medium and long-lived secondary species which start with zero concentration in the start and must build up on the first run.

Simply put   the spin up time is the time taken by the model to get balanced between the mass and velocity fields in the initial period of integration. 

Because it is not easy to illustrate my answer without equation, I put a link with some summary.

1: Lots of simple models can run without spin-up because direct solver (MUMPS, etc.) and observations can be used to solve the ODE/PDE directly if you have the initial condition and boundary condition; and they can take a long time to do the matrix operations;

2: When we cannot solve the equation directly, we can use iterative methods. For example, lots of models use PETSc to find a solution.

3: Another method often used is the spin-up, it can be used to approximate the state variables; it can also take a long time because the pool and fluxes are not in the same order;

4: With spin-up, we also need to be careful with parameters and threshold because of local/global optimization issue.

5: Mathematically, spin-up is similar to forward explicit/Euler finite difference method whereas iterative matrix solvers are implicit finite difference methods.

https://changliao.blogspot.com/2017/09/numerical-simulation-001.html