I am evaluating the thermal conditions of passengers inside a car cabin using this SimScale tutorial: Advanced Tutorial: Internal Car Thermal Comfort.
The CAD model in that tutorial aligns with my simulation requirements, but I have questions regarding the simulation setup and mesh definition:
To train an ML model to accurately assess the thermal comfort of passengers (PMV), I need to simulate different environments and collect the data from each simulation. Would a steady-state or transient simulation make more sense for me? I need at least 10,000 data sample points, preferably from 10+ simulation scenarios.
My previous simulations have either failed because of divergence, or completed but with divergent results that are extremely faulty. I only care about the temperature metric in the results field. Is my divergence caused by the mesh definition or the CAD model? I think the model itself has some very small areas that are causing this divergence but this is the same model that was successfully simulated in the tutorial so I find that very strange.
How many simulations can I concurrently run on the same SimScale project without affecting the computational time of each individual simulation? That is to say, If one simulation takes 2 hours to complete normally, would it need more time to complete if I run a parallel simulation?
It is not cost-efficient to train ML models using data from transient simulations unless the focus of the training (and objective functions) is related to transients (for example, the time required to cool the cabin) or transient phenomena (such as flow separation). Furthermore, standard ML models will struggle to predict time-dependent variables, as some behaviors cannot be related to changes in other properties.
If your focus is thermal comfort once the thermal balance within the car has reached a steady state, you should use a steady-state model.
You can boost parallel simulations using a cost-effective approach by using 1 processor for every 70k to 130k mesh elements. So, if your total element count is, for example, 600k, don’t use more than 6-8 processors.
