I am trying to analyze the sensitivity of different parameters towards the snow simulation. Can anyone suggest which methods is best for doing so?
Design of Experiments is the first you should try, e.g. http://uk.mathworks.com/help/stats/design-of-experiments-1.html
Your choice of SA method depends on how many parameters you would like to include in the SA and on the computational time required for simulating your model (and on the time available). If you would like to carry out an uncertainty quantification, the choice of method also depends on information available on input (parameter) uncertainties. Provide a rough estimation on the above, and then I can try to suggest a method.
If your numerical forward model has been validated, then it depends on the way you run the forward model. Generally all the SA methods are available. But variance based (VB) method requires intensive evaluation of the forward model, then the meta-model is necessary to replace the forward model; while element effect (EE) method is also powerful even with limited evaluation of the forward model.
Dear Magnus, i have around 10 parameters. I am planning to run the model for data of one year with daily time step.
Dear Amrit
Global sensitivity analysis (Sobol' indices) is definitely the most popular method these days. In conjunction with surrogate models such as polynomial chaos expansions (PCE), this is probably the most efficient method around.
If your model has 10 input parameters, about 100-200 runs (eg, Latin Hypercube Sampling) should be sufficient to build a PCE, from which you get analytically the Sobol' sensitivity indices.
Literature (preprints also exist)
http://www.sciencedirect.com/science/article/pii/S0951832007001329
http://www.sciencedirect.com/science/article/pii/S0951832010001493
All these tools are available in the recently released platform UQLab: http://www.uqlab.com/ which is free of use for research.
(see the manual about sensitivity: http://media.wix.com/ugd/0f4bf0_e0ea2f28dbac414faf64c9493a610b97.pdf)
Hope it is useful!
Dear Bruno,
thank you for your suggestions and solutions. Will try it and let you know.
- Badges
- Science topic
- Similar topics
- Methodology
- Research Methodology
- Research Papers