Our website uses cookies. By using the website you agree ot its use. More information can be found in our privacy policy.

LS-OPT 3.4 released

LS-OPT Release 3.4 -- Informational Document


Since Reliability-based Design Optimization (RBDO) is rapidly becoming the standard for simulation-based design optimization, LS-OPT Version 3.4 focuses on several improved features for this design application. Significantly faster computation has been achieved by improving the reliability analysis code, introducing new optimization algorithms and optimizing the executable code. User-friendliness for both optimization and probabilistic analysis has also been improved, principally by adding a completely new wizard for analyzing LS-DYNA statistics and simplifying the choices for metamodel-based optimization. Several Viewer (post-processor) features, among others the visualization of the Pareto Optimal Frontier, have also been enhanced.

In the sections that follow, the main new features are discussed, principally

  • (i) simplified optimization strategies,
  • (ii) faster optimization algorithms,
  • (iii) LS-Dyna statistics wizard and
  • (iv) enhanced visualization of the Pareto Optimal Front.

Strategies for Meta-Model based Optimization

The purpose of the new "Strategy" panel is to provide a simple choice of basic optimization setups depending on the application, while avoiding the pitfalls. There are three recommended strategies for automating the metamodel-based optimization procedure. Determining their availability is (i) whether the user wants to create a surrogate design model for design exploration (such as the creation of a Pareto Optimal Front) or (ii) whether he/she is interested in finding a single improved or optimal variable set (as e.g. for material parameter identification).

The different strategies are shown schematically in Figure 1. They apply to all metamodel-based optimization tasks, including RBDO and were possible in Version 3.3 but the setup was more complicated. Selecting a strategy in the new tab will change and simplify the available options in subsequent steps such as sampling, algorithm selection and running.

Fig1: Metamodel-based optimization strategies in LS-OPT

Fig1: Metamodel-based optimization strategies in LS-OPT

The first selection is whether a Pareto Optimal Frontier is required. Selection of the POF then limits the available options to the global strategies: Single Stage and Sequential. The remaining option (Sequential with Domain Reduction) is typically only used for optimization in which the user is only interested in the final optimum point (such as parameter identification) and not in any global exploration of the design.

Meta-Model Optimizers

Until Version 3.4 the two optimizers available for metamodel optimization were the gradient-based algorithm LFOPC (Leapfrog Optimizer for Constrained problems) and the Genetic Algorithm (GA). The LFOPC algorithm is highly accurate and robust but can be time consuming for large optimization problems, especially when using RBDO. This deficiency is largely due to the fact that LFOPC is a local optimizer and therefore requires a multi-start approach in an attempt to find a global optimum. The GA was implemented for Version 3.3 to address multi-objective constrained optimization problems but also serves as a global optimization technique. Adaptive Simulated Annealing (ASA) has now been added to address the global optimization problem more efficiently. To further improve the accuracy of global optimization, a hybrid approach has been adopted. In this approach, a global optimizer is used with limited computational budget to drive the search to the global optimal region. Next, the sub-optimal solution from the global optimizer is used as the starting point for LFOPC to sharply converge to the global optimal solution.

LS-DYNA Statistics Wizard

LS-OPT can display statistical results visually in LS-PREPOST on the structure. This ability has been completely redesigned to allow:

A shorter learning curve. The GUI wizard and outlay guides a user through the creation of a plot.

Increased usability. The capability was re-organized to focus on plots as the central entity. These plots can be edited, displayed, computed in a batch fashion, and refined further (for example, in a bifurcation analysis).

Re-use and sharing of an investigation method. The plots created during the analysis of a structure is saved in a database. This database can be re-used in similar studies. For example, a metal forming group have to set up their investigation methodology only once and re-use this set-up for a number of similar metal forming studies.

Fig2: Standard deviation of plastic yield of a metalforming problem. The creation of plots of statistical quantities like these has been simplified by the new GUI.



Fig2: Standard deviation of plastic yield of a metalforming problem. The creation of plots of statistical quantities like these has been simplified by the new GUI.

Visualization of Pareto Optimal Front

Visualizing the Pareto Optimal Front is important for choosing an appropriate design from the set of optimal designs. Displaying the Pareto optimal front for a problem with only 2 objectives is simple. However, for more objectives, more sophisticated display features are required. Two approaches are taken to visualize the Pareto Optimal Front. The first is to allow the display for problems with up to 4 objectives and has been implemented in Ver. 3.4 (see Figure 3 for a 4-dimensional display). The fourth quantity is displayed using a color index.

Higher dimensions require more sophisticated methods for mapping the multiple dimensions to a 2D display for selection of an appropriate optimal design. Three such methods, each with its own favorable attributes, are being implemented in Version 4.

Fig3: 4-Dimensional visualization of the Pareto Optimal Front for triple objective crash optimization problem: HIC (Head Injury Criterion) vs. Intrusion vs. Mass with a thickness parameter shown in color (LS-OPT GUI)


Fig3: 4-Dimensional visualization of the Pareto Optimal Front for triple objective crash optimization problem: HIC (Head Injury Criterion) vs. Intrusion vs. Mass with a thickness parameter shown in color (LS-OPT GUI)

Miscellaneous other Features

  1. A feature has been added to evaluate design points using an existing metamodel. A .csv file containing all the interpolated design results is produced. The feature can be selected under the "Evaluate Metamodel" tab in the "Solvers" panel. The repair feature "Analyze checkpoints" is used to do the evaluation using an existing database.
  2. The methodology for sampling within a reasonable design space has been improved to make sampling constrained by geometric or other limitations more robust (so called "move" option). This feature is now also available for the direct GA optimizer.
  3. More attributes are provided in the "Accuracy" plots such as clickable points to show computed/predicted values, and the feasibility status.
  4. Most databases are now also available as a .csv (comma separated variables) file for importing into spreadsheet programs such as Microsoft Excel (migrated to Version 3.3).
  5. Discrete sampling is also available for the Space Filling sampling scheme.
  6. The detailed optimizer history is available for each algorithm (OptimizerHistory_n.csv). These are available as .csv files, but will also be displayed in Version 4.
  7. The LS-OPT database gathering feature (.zip file) has been extended to include the history data for each simulation run. These are required for the DynaStats and MeanSqErr features.
  8. The summary report (lsopt_report) files have been extended to all the tasks.
  9. The Kriging metamodel has been updated for better speed performance.

Closure and Outlook

LS-OPT Version 3.4 presents a significant step forward for design with LS-DYNA by providing a friendlier interface as well as refining and speeding up the existing methods for probabilistic analysis and optimization. New post-processing features such as extended visualization of the Pareto Optimal Front have been added.

Version 4 is being developed with a new generation LS-OPT post-processor which, in addition to the current features, have several new data mining features focusing on variable/response correlation, visualization of the Pareto Optimal Frontier and response history visualization. A standalone tool for nonlinear topology optimization has also  been developed for release in 2009 (2nd quarter).