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

Release of LS-OPT® Version 4.2

Livermore Software Technology Corporation is pleased to announce the release of LSOPT ® Version 4.2. A major focus of the V4.2 development has been to refine and enhance the LS-DYNA® job distribution features for running LS-OPT on a PC or Linux
machine controlling and monitoring LS-DYNA jobs distributed on a Linux cluster.

In a further major development, the material parameter identification feature in LS-OPT has been enhanced using a new Curve Mapping feature to enable complex curve matching typically used for the calibration of highly nonlinear hysteretic models.

Several other important features are summarized in the following overview.

1. Curve Mapping for Parameter Identification

System parameter identification is a prominent feature of LS-OPT®, typically used for the purpose of calibrating material models. The procedure consists of minimizing the mismatch between two curves, of which one represents an experimental result. In addition to the optimization algorithm, a curve matching metric, which quantifies the mismatch, is a critical part of the methodology.
In Version 4.2, a new and simple curve matching approach has been introduced. The purpose of the approach has been to address the following problems typically associated with curve matching:

  • The output is represented by a hysteretic curve which may have multiple ordinate values (e.g. force) for a particular abscissa (e.g. deformation). Hysteretic curves are common in materials testing.
  • Less severe than the above, but still potentially problematic, is the possibility of encountering non-hysteretic curves but with very steep sections such as typically occur in damage models.
  • Comparative curves are typically of different lengths precluding the use of full curve lengths for the mapping.

To address material or system calibration problems in general, a Partial Curve Mapping method has been introduced in LS-OPT. The feature involves the ability to continuously map one curve to another without having to be concerned about the relative curve lengths or shapes. The LS-OPT GUI simply requires the names of the experimental and test curves without any additional attributes. An example result is shown in Figure 1.


Figure 1: An example of the calibration of a material model with hysteretic behavior using Partial Curve Mapping. Curves at several iterations are shown in the upper plot. The black crosses represent the experimental target curve. The optimization history of the curve match discrepancy appears below.

2. LS-DYNA job distribution and the LSTCVM Secure Proxy Server for clusters

A popular execution mode for optimization is to run LS-OPT on Windows while launching, controlling and monitoring LS-DYNA on a Linux cluster. For this purpose, LSTCVM was created as a secure proxy server to allow on-line job monitoring in a
secure network-based computing environment. LSTCVM allows the administrator to set up restrictions regarding allowable commands, run locations, users and interactivity. No login is required and hence no passwords are transmitted. LSTCVM interfaces with all the queuing systems currently available in LS-OPT.

To address the problem of individual straggling solver jobs, i.e. jobs of which the run times become unexpectedly long or jobs lost in the queuing system (this happens occasionally!), a job cancelling feature has been added for individual jobs. This feature allows the user to cancel a particular simulation to avoid holding up the entire optimization run. Normal and accelerated cancellation features are available. The ability to cancel individual jobs is in addition to the existing feature for stopping the entire optimization job. The loss of a few individual jobs is usually not critical for the overall optimization result since LS-OPT provides some redundancy in this respect.

Along with these specific features, the job distribution facility has been significantly refined and tested in industrial environments involving very large clusters. As in the past, job progress can be fully monitored and the viewing of any solver job log requires a single mouse click.

LS-OPT and LS-TASC utilize the same job distribution system.

Figure 2: LS-OPT job progress display showing list of running jobs including a single cancelled job. A specific job running on a cluster can be viewed or cancelled by selecting the desired job button under the View Log.

3. Progress visualization and stopping criteria for Direct Multi-Objective Optimization

After Multi-Objective Optimization was introduced in LS-OPT in Version 3, it became apparent that unless an accuracy-based termination criterion was provided, this type of analysis could potentially be unnecessarily expensive. Direct Global Optimization
algorithms typically require a large number of simulations to obtain accurate results and should therefore only be run to a specified level of accuracy.

Version 4.2 provides a number of stopping criteria of which the Crowding Distance (representing optimal design point density) and the Dominated Hypervolume are the most intuitive to use. The Dominated Hypervolume is defined as the volume between the optimal solutions and a reference point in the objective space, so in 2D is simply the surface area above the front (see Figure 3, top left).

For deeper insight into the optimization process, the histories of various attributes of the POF can be plotted. These attributes include: Archive size, design points removed or added or carried over, hypervolume (and its change), crowding distance, spread and minimum and maximum objective values. Some of these are shown in Figure 3.

Figure 3: Result of a Multi-Objective Direct Optimization problem showing the convergence history of the crowding distance (standard deviation) (top right) , archive size (bottom left) and dominated hypervolume (bottom right) of the Pareto
Optimal Front displayed at the top, left. Various criteria, including the hypervolume change can be used to terminate the optimization at a prescribed level of accuracy.

4. Kinematics for crashworthiness

Safety requirements related to crashworthiness design often require the computation of distances between nodes or relative displacements (i.e. deformations) of a crashed
structure. For this purpose deformation and distance have been added as output quantities that can be extracted from the LS-DYNA nodout database.

The feature makes it much easier to compute quantities in global or local coordinates and now makes it possible in the reference coordinates of a predefined moving rigid body.

Both histories and responses can be extracted.

Figure 4: LS-OPT Interface for nodout quantities showing dialogs for specifying local deformation relative to a rigid body fixed to the structure. The rigid body definition requires the specification of three nodes on the finite element model.

5. Design sampling using Constrained Space Filling

In many design problems constraints exist on the relationships between design parameters, so that arbitrary parameter combinations do not necessarily represent analyzable designs. Such examples are typically shape design problems in which there is an interdependency between the parameters. The design space representing only the possible designs is often referred to as a reasonable design space. As a simple example, two adjacent radii of an opening are dependent on one another so that their combined radii cannot exceed the size of the hole. This type of limitation results in an irregular shape of the design space which complicates the sampling method employed to build metamodels of the design.

In Version 4.2, Constrained Space Filling has been introduced as an improved method for providing sampling points within the reasonable design space. A 2-dimensional example is shown in the Figure below.

Figure 5: LS-OPT display of an example with points sampled in a 2D irregular design space. The method used is a Constrained Space Filling approach.

6. Selected other new features

  1. Injury Criteria for crash design. A comprehensive set of injury criteria for crash analysis is now available, the total list being HIC, NNIC, NIC, Nkm, LNLI, Chest Compression, Viscous Criterion, TTI, Chest Severity Index, Tibia Index, a3ms (3ms acceleration threshold) and Clip3m. A 3-node version of the injury criterion Clip3m has been added. Some criteria are available as histories.
  2. Stopping criteria for sequential metamodel-based optimization. Metamodel prediction accuracy based on the PRESS error has been added as a stopping criterion for the classical Sequential Response Surface Method (SRSM). This feature can also be used as a stopping criterion for Multi-Objective Optimization problems based on sequential metamodel updating. The PRESS error defines the prediction accuracy of the metamodel, so allows the optimization to stop when the accuracy no further improves.
  3. Viewer refinements. The display of LS-DYNA histories and crossplots was improved by allowing the selection of multiple histories. This makes it easier to display multiple history vs. test comparisons in multi-case problems. The 2D Metamodel Interpolator display now also includes simulation points. In general, homogeneity amongst the different types of displays has been improved and all metamodel-based displays have been significantly accelerated.
  4. Generic Extraction from text files. Histories have been added to the GenEx (generic extraction) result extraction feature. GenEx allows the extraction of quantities from text output files by locating user-defined markers at desired locations in the files. The feature allows LS-OPT to interface to non-LS-DYNA solvers. In earlier versions, GenEx could only be used to extract responses.
  5. LS-OPT database archiving has been expanded to optionally include extra files such as solver input files.