Table of Contents (click for easy navigation)
- Overview
- How to conduct DOE in ANSYS software?
- Defining objectives and parameters
- Choosing a Design exploration method
- Choosing a DOE method
- Creating a response surface
- Common issues with response surface and DOE
- DOE optimization algorithms in ANSYS software
- References
Overview
Design of Experiment (DOE) is used to sample a design space which includes all the design parameters so that a statistical model can be built in order to predict responses. Responses such as maximum and minimum stress, the maximum and minimum temperature and first natural frequency can all be determined by DOE. DOE is specifically useful when the user can sample a limited number of points. With the DOE in place, the users can create complex geometries to be tested in a virtual world (ANSYS Inc., 2018).
The main purpose of Design of the experiment is to spread out the samples so that the results have a low uncertainty. This also results in high accuracy and precision in prediction.
How to Conduct DOE in ANSYS Software?
Defining objectives and parameters
In order to carry out the design of the experiment for a given model, the objectives and list of design variables must be defined first. To do so, open the “project schematic” window which should look like this;
Choosing a Design Exploration Method
In the design exploration window, select the “Response surface.” This allows the users to carry out the DOE to create a predictive model known as the response surface. Drag the “response surface” tab from the toolbox in the project schematics.
Choosing a DOE Method
ANSYS software comes with many different DOE methods that can be used for testing purposes. Two of the commonly used DOE are Optimal space filling with user defined sample points and LHS (Latin Hypercube sampling) (CAE Associates, 2017). One of the biggest advantages of these methods is that the number of parameters and the number of samples is completely independent of each other giving the users more freedom to carry out the DOE. The Central Composite Design (CCD) is not recommended to be used by beginners as in many cases CCD objective cannot accurately calculate via a quadratic function and requires advanced inputs by the users. CCD also has a long processing time and requires high processing power (Cunningham, 2015).
Click over the “Design of Experiments type” and select the desired design type and DOE;
Double-clicking the DOE will open a new window where all the inputs and outputs parameters can be set.
After the parameters are set, click ‘preview’ to see the lists of DOE according to the settings;
Now select ‘update’ and wait for the processing to take place. The processing time varies from hours to several days depending on the samples you have and the processing power of your computer. However, it must be noted that in some cases, the DOE samples cannot be fully evaluated due to any significant changes in the geometry.
Creating a Response Surface
The results obtained from DOE can be used to generate response surfaces which further can be used for prediction purposes. ANSYS software has the following list of response surface methods;
- Standard response surface
Standard response surface makes use of polynomial surface to fit the data. In terms of computing power for prediction and fitting, standard response surfaces use the least amount of power. This method of surface response is recommended only if you have lots of data to work with and there is a smooth change in objective.
- Kriging
Kriging is a non-parametric type of surface response, and its prediction is dependent over the existing data points. In cases where there is a large number of data to work with, kriging tends to be a slower process. Though this method might be slow in some instances, kriging can automatically fit through all the data points. Kriging is recommended to be used when your data is limited and non-linear, and you completely trust your simulation results to be accurate.
- Neural Network
It creates a nonlinear mapping from the product design (input) to the objective value (output) that mimics the expensive simulation. The neural network is slow in training and can take quite some time to establish; however, once set up; it can generate accurate results. This method is recommended to be used when your data is huge and is highly non-linear.
- Non-parametric regression
Non-parametric regression makes use of support vector regression and is quite like kriging. However, instead of using all the data, non-parametric regression determines the most important data points and perform its predictions. Therefore, the computational power for this method to work is low, but the cost of fitting is still high. This method is recommended to be used when the data is highly non-linear, and you do not require your model to fit right through your data.
Common Issues with Response Surface and DOE
Doing your DOE and creating a response surface demands a lot of trials and errors for someone who is not much experienced in it. In order to counter those issues, here are some of the solutions proposed;
Issue | Probable reason | Solution |
My verification results are bad | Your sample size is small. Response highly non-linear. The surface response is too flexible resulting in overfit | Increase the sample size. Try a different model for data fitting |
My mode has a bad fit | Small sample size. Low complexity of the response surface | Increase the sample size in DOE. Increase the complexity of the response surface |
My mode takes days to process the results | Your sample size is huge. You are using a slower model to compute. Your computing powers are limited | Reduce the sample size if possible. Use faster DOE models. Make your computer can carry out high power processes. |
I cannot determine which model is the best | You are unable to use the ‘judge/verification point’ command. All the collected data was used for training the model | Collect some data aside for training the model. Use the validation error |
DOE Optimization Algorithms in ANSYS Software
ANSYS software provides the user with several different optimization algorithms, some of which are discussed below;
- Screening
Screening is the random sampling of the space to pick out the good sample from the data. Screening can be used to give the user an initial trial of the setup. For example, to determine whether your simulations give reasonable results or not, screening is an ideal algorithm to use.
- Adaptive single-objective optimization (ASO)
ASO makes use of optimal space filling for the design of the experiment, the kriging as the response surface and the MISQP for the determining of local optimal solutions from the response surface. ASO is used when the constraints and objectives are expensive, and there are limited budget and time for optimization.
- Multi-objective genetic algorithm (MOGA)
MOGA is one of the most commonly used algorithms and is used to determine the Pareto-optimal designs simultaneously. MOGA is used when there are multiple objectives in the solution.
- Nonlinear programming by quadratic Lagrangian (NLPQL)
NLPQL is a fast-local search algorithm. It is primarily used when the simulation time is in minutes, when there is only one objective and when the number of variables is small (<10).
- Mixed Integer Sequential quadratic programming (MISQP)
MISQP is similar to the NLPQL, but it also allows for the integer variables. MISQP takes more processing time as compared to NLPQL due to the addition of integer variables.
DOE and design optimizations are one of the crucial phases in industrial grade product designing and prototyping. For product designers to not come up with flawed products off the assembly line, it must be made sure that the DOE is done right and the design is thoroughly optimized.
References
- ANSYS Inc., 2018.
ANSYS Mechanical Enterprise-Design of Experiment. [Online]
Available at: https://www.ansys.com/products/structures/ansys-mechanical-enterprise
[Accessed 15th March 2019]. - CAE Associates,
2017. Virtual prototyping offers enhanced product innovation. [Online]
Available at: https://caeai.com/ansys-software-support/ansys-benefits
[Accessed 10th March 2019]. - Cunningham, P.,
2015. How Does my DOE Model Determine How Many Design Points to Generate?. [Online]
Available at: https://caeai.com/blog/how-does-my-doe-model-determine-how-many-design-points-generate
[Accessed 10th March 2019].