We all feel the “need for speed” when trying to find better design solutions through simulation. What can we do to speed up our design exploration studies? Let’s discuss all the options here, including one that may not be obvious.
In a typical study, the total CPU time needed to perform a design exploration study is determined using this simple formula:
This formula contains three factors that determine the time needed to execute a study:
Number of Evaluations
This is the number of different designs (or combinations of design variable values) that will be evaluated during the search process. We often refer to this as the “exploration budget.”
Because the required number of evaluations depends on the complexity of the design space, we could speed the search by reducing the number of design variables and/or the range of the design variables. Each of these approaches has the potential to simplify the design landscape, making it easier for the search to converge. However, this simplified design landscape might not contain the designs you hope to find or that new concept you did not know existed. So limiting the amount of design exploration might speed convergence, but it might also result in lower final design performance. As a general rule, we suggest defining the problem using your best judgment of what variables and ranges are most likely to yield the best designs. If you are mostly interested in refining a particular design, then a smaller number of design variables or ranges might make sense.
Of course, the best way to reduce the number of evaluations across all problems is to use an efficient search strategy like SHERPA. This hybrid and adaptive search approach works well on a wide range of problems, from very simple to the most complex.
Average Time per Evaluation
Almost all of the CPU time spent during a design exploration study is associated with performing simulations of designs. The execution times of these models can often be shortened by reducing the resolution (e.g., using a coarser finite element mesh/grid) or by simplifying the model representation (e.g., de-featuring the model or using sub-models). Creating more efficient models is usually a good idea, but not if it means sacrificing accuracy. An inaccurate model can mislead the design search, making the final result an “artifact” of the model instead of a truly improved design. There is a quote attributed to Einstein that says a model should be “as simple as possible, but no simpler.” Excellent advice.
If multiple independent simulations need to be performed for each design variant (e.g. crash and NVH), running these simulations simultaneously in parallel affords another avenue for search acceleration.
Additionally, we often have the ability to run a single simulation on multiple processors, sometimes approaching nearly linear speedup for certain levels of parallelization. When the necessary computer hardware and software licenses are available to make this possible, parallelization is a great way to reduce the clock time for an exploration study. These days, clock time is often more valuable than CPU time.
Number of Parallel Evaluations
Speaking of parallelization, some search algorithms (like SHERPA) have the ability to work with multiple design evaluation results at a time. In other words, instead of deciding where to search next after each single design evaluation, SHERPA is able to decide which N designs to evaluate next. It then uses the results from these N designs to decide the next set of N designs to evaluate, and so on. In this case, linear speedup is achieved, so the study can be run up to N times faster.
There are practical limits on the size of N, however. SHERPA needs to analyze past results periodically so that it can redirect the search and use future evaluations in the most efficient way possible. As a rule of thumb, it is recommended that N be no larger than the total number of evaluations (the exploration budget) divided by 10. This will give SHERPA at least 10 opportunities to analyze previous results and adjust the search strategy. More than 10 is preferred.
Bonus Speedup Factor – Collaboration
There is one more solution time factor that is not included in the above formula. In fact, it just might be the most important factor of all – intuition. Intuition plays a critical role before, during, and after any design exploration search, including at the beginning of a restart. Some aspects of intuition are included within the above three factors, but there are more direct ways to make use of it, such as injecting actual designs before and during a study. We call this collaborative design exploration because it couples highly developed human intuition with advanced mathematical and computational capabilities. Given a sufficient exploration budget, HEEDS will be able to identify better performing designs in the design space on its own. But, by coupling its specialized search capabilities with fundamental domain knowledge (through design injection), high performing regions can be explored more quickly and effectively, leading to truly innovative breakthroughs. A full coverage of this topic is outside the scope of the current post, but we will discuss it more in the future.
We hope these tips helps you to discover better designs, faster.