Validation of simulation results

All models are wrong, but some are useful. That phrase has always stuck in my head for some reason. It reminds me that the assumptions we make when modelling real systems are always deviating from true reality. Although models are always missing some piece of reality, they can still be excellent predictive tools within certain bounds.

Simulations in isolation have fairly limited utility in my opinion. It is only in the context of experimental data that the simulation can become useful.

Parameterising the model with experimental data is one thing, but to have confidence about the limits for which that model is accurate is another. I tend to have the approach of: use the simplest model that does the job. Anything more complex is unnecessary and can lead to potential overfitting of parameters, particularly in so-called "sloppy" models.

A good model:

• Has accurate predictive power outside of the dataset used to parameterise it
• Includes the minimum number of parameters required to capture the system behaviour (Occam's razor)
• Offers insight into the physics being modelled - i.e. not a black-box