Development teams doing simulation-based engineering can boost their productivity using Design of Experiments (DOE) performed early in the process – planning virtual experiments to deliver a maximum of relevant design insights at minimum simulation cost. Now they can go one step further with Optimus Adaptive DOE to actively exploit the design space knowledge being built up as the experiments run.

This automated active learning process adds extra virtual experiments where needed most in order to better capture the design space. In addition, it delivers a Response Surface Model (RSM) that more accurately condenses even the most complex physics simulation models into an inexpensive-to-use surrogate model. Obviously such RSMs provide an excellent foundation for ultra-fast design optimization. In addition, these models can readily be exchanged with and integrated into any environment for model based system development using the Functional Mock-up Interface (FMI) open standard.

DOE delivers focus on what matters most

Development teams performing simulation-based engineering still tend to follow a sequential trial-and-error approach to gradually improve product design. First the engineers build an initial virtual model, run physics simulations and interpret the results, and then iteratively test a number of design variants they think might deliver improved product performance.

These engineering teams can boost their productivity by having Optimus perform Design of Experiments (DOE) early in the process. Based on a set of well-chosen virtual experiments to sample the design space, they can easily identify the most influential design parameters and discover the relevance of design constraints. This information is critical for them to focus on where the biggest potential is to optimize the performance of the design.

DOE enables fast design space exploration

Using Optimus, simulation results delivered by DOE methods can be condensed into surrogate models or RSMs. In particular when the underlying physics simulations are computationally expensive, these RSMs enable fast design space exploration & optimization (in seconds rather than days or weeks).

A drawback of this approach is that the entire DOE sample set is defined using existing information before any physics simulation results become available. This ‘black box’ approach makes it hard to know in advance which and how many samples are required to achieve a given quality level for the surrogate models that will be built on it - bearing the risk of undersampling the most relevant design space regions (or oversampling less relevant regions).

Active design space learning through Adaptive DOE

Adaptive DOE removes the hurdles of this ‘black box’ DOE approach. Adaptive DOE (also known as sequential DOE) analyzes the available DOE data points to complete them with extra samples in design space regions of interest.

Starting from an initial DOE design, typically a Latin Hypercube method, RSMs are built for the experiment responses of interest. Optimus will then add extra DOE samples where needed, update the RSMs, and iteratively continue to do so to deliver models that are best fit for purpose – accumulating design space knowledge as the process runs. By adding DOE samples in batches, the user can benefit from Optimus parallelization capabilities to speed up the process. Optimus offers the flexibility to stop the process at any time, for instance when the required accuracy is reached and/or when the allocated simulation budget is spent.

From design space exploration to exploitation

Optimus enables the engineers to flexibly define the strategy of the Adaptive DOE procedure by putting more emphasis on design space exploration than on exploitation, or vice versa.

Emphasizing Exploration will direct Adaptive DOE to expand an existing DOE dataset and deliver best overall coverage across the design space (including border areas) through a more uniform sample spread. This strategy allows engineers make better use of existing DOE data.

Instead, focusing on Exploitation will add new samples in regions of particular interest by taking into account model responses – either starting from an existing DOE dataset or from scratch. By improving the accuracy of surrogate models in regions that are either very interesting (local optima) or difficult to model (discontinuities, steep slopes, etc.), the engineers achieve more accurate modeling of complex (and in particular non-linear) design behavior.

Efficient modeling of complex physics behavior

Optimus Adaptive DOE delivers high-quality RSM models for even the most complex physics behavior. Performing design optimization directly on such RSM models avoids computationally intensive physics simulations altogether, tremendously saves on simulation time and uses available hardware and software resources as effectively as possible.

The obtained high-quality RSM models can also be exported as Functional Mock-up Units (FMU), condensing detailed complex physics models into an accurate and inexpensive-to-use equivalent. Using the Functional Mock-up Interface (FMI) open standard, RSM models can easily be exchanged with and integrated into any environment for model-based system development.

Want to learn more about Optimus’ Adaptive DOE implementation and how it can be applied to your design space exploitation challenges? Send us an email ([email protected])!