Exploratory versus final optimization analyses

To determine the best settings for your thermal correlation analysis, it is recommended that you start with exploratory analyses before you run your final optimization analysis.

When you correlate and optimize a thermal model, a first exploratory thermal correlation analysis is useful to get more information about:

Sensor values that are the most difficult to reach.
Design variables that influence the most and the least the correlation.
Correlation times that you should specify for the final run.
Convergence criterion to use.

Use the results from the exploratory runs and the guidelines for TMG Correlation settings described in the following table to set up, optimize, and update your original thermal solution.


Settings	Exploratory analysis	Final optimization analysis
Targets	Do not use weights except if you already know which sensors are more important.	Add weights to your sensors to focus more on these weighted sensors when reducing the gap with the test data. For example, when a sensor is located near a critical component of the model, more weight on this sensor makes TMG Correlation focus more on reducing the gap with reference data at this sensor location.
Targets	Note: Adding weights to sensors can make your optimization problem stiffer.
Optimization design variables	Always specify minimum, maximum, and initial values for all optimization design variables. The minimum and maximum values define the design space for the variable. The wider the range between these values, the larger the design space becomes, allowing for more possible solutions. The initial value defines from where the optimizer starts minimizing the objective function. A different initial value may lead to a different solution.
Correlation times	By default, TMG Correlation lists all sensors time points and solution time steps. Considering all these correlation times is similar to optimizing the area between the sensor and solution curves. Note: When the optimizer has more correlation times, it has more targets to reach, thus, the optimization problem is more complex.
Correlation times	Use few correlation times as follows: Avoid using all sensor time points. Select times that have the highest differences between simulation and reference data. For quasi-steady state sections, use only one additional time within the section start and end times. For transient sections with steep slopes, use at least one additional time within the section start and end times.	Using the exploratory analysis results, you can identify the times where the temperature difference is still large and add one or two correlation times to improve the correlation.
Convergence criterion	Use a constant convergence criterion that satisfies your accuracy expectations for all correlation times.	If you have different accuracy expectation for different parts of your transient problems, use the time varying convergence criterion option to specify different convergence values for the different correlation times of your model, thus, satisfying your accuracy expectations for every correlation time separately.
Convergence criterion	In both runs, estimate the initial objective function value using the initial gap between simulation and reference data. Calculate the average absolute error for every correlation time of your problem. This gives you an idea how the objective function and average absolute errors are linked. Suggested values for convergence criteria: For a maximum error of 5 degrees with the reference data, define a convergence criterion of 1. For a maximum error of 0.5 degrees with the reference data, define a convergence criterion of 0.1.
Maximum design cycles	Fewer than 5 design cycles	10 to 20 design cycles are usually necessary to converge to an optimal solution for a well-defined correlation setup depending on the complexity of your correlation problem. If this is not enough, increase the value for your next correlation.
Optimization algorithm	If you have uncertainties about the ranges (min, max) of your design variables, use the default optimizer, which is MMA. If you are confident about the ranges (min, max) of your design variables, use the SLSQP optimizer.	When the min and max values of the optimization design variables are defined, SLSQP performs better and faster than the other algorithms.