ODYSSEE Lunar – GUI
- In the dialog box 4-interpolation, the SVM method with option classification is now the new SVM function proposed in ODYSSEE Solver 2023.1. For this implementation, the argument list in the dialog box has been updated to select the kernel type and chose the C-value, Gama value and degree value.
- In the optimization part, the optimizer name NLPQLP (Local + constraint) and SQP2 (Local + no constraints) are respectively renamed NLPQL and SQP.
- Update of the generate range tab in 3-Sensitivity, allowing changes to the min / max range of the parameter to vary. This gives a shorter corridor to test.
- In the animation tab, for the d3plot files prediction, it is now possible to see the translational displacements as it was done for the h5 files.
- Updated the dyna_output file by adding the rotational speed and rotational acceleration extracted from the ascii nodout file to the output list of the LS-Dyna analysis tool.
- Updated the Lunar Help document.
ODYSSEE Quasar – Solver
- A Quasar-Python wrapper has been developed to use the Quasar’s external functions in Python code. For details “How to use it?”, please refer to Python Wrapper chapter in the Quasar help.
- Solver:
- A new solver is available in Quasar, the SVM (Support Vector Machine) option classification dedicated to classification problems using hyperplanes to separate data.
SVM("classification", X, Y, XN, kernel_type, C, gamma, degree, save_files, prefix)
- Backward compatibility to run POD binaries between 2022.3 and 2023.1.
- Quasar Help document:
- An LSTM example has been added.
- SVM example has been added.
- Two Useful functions have been added:
- cumtrapz (cumulative trapezoidal integration) function computes an approximation of the cumulative integral of Y via the trapezoidal method with unit spacing.
Matrix cumtrapz(Matrix Mat, Matrix IntegrationAxis, int orientation)
- rolling_ave function calculates the rolling average on a full curve. The mean is based on the previous points located in a window where the size is given.
Matrix rolling_ave(Matrix Mat,int windowSize,int orientation
- FMU Improvements
- At the beginning of the xml file, a xml tag <fmiModelDescription> is added containing the version of ODYSSEE CAE used for the generation, the creation date and the current version of the FMU.
- Updated *.xml <ModelStructure> at the end of the xml file by adding <InitialUnknown…> information, thus fixing some execution errors in ADAMS.
- Improved performance for prediction in FMU ROM Smart Superelement (POD_krg mainly).
It improves performance for large training datasets. - Still in the performance improvement, the POD solver does not require Y datasets anymore in the FMU *.zip file and will not be saved in it starting in 2023.1. It helps to reduce the FMU’s memory usage.
- Some fixes have been done on the *.qsr file inside the FMU for the data loading in memory.
- The FMU generated in Windows OS through ODYSSEE CAE are compatible with an execution with ODYSSEE Solver Linux Redhat 7.9.
- Bug Fixes
- For the prediction of the animations with the option “elemental data” selected, when POD was selected with “recompute database file” unselected (save_file=-2) the calculation took the same time as the option “recompute database file” (save_files= -1).
- For the curves prediction, with mode number different of 0 (0=all modes), the result was incorrect.
- For the INVD predictive method, if the new point to be predicted (Xn) was already known in the X-learning database (Xn=X), then the result Yn obtained was from the predictive formula. We changed it to give a result equal to the known value in the Y-learning database.
- The Tx argument in LSTM function (length of one sequence of data = number of points describing one repetitive cycle) didn’t work.
- The function order_y_by_row_index_x has been fixed. It correctly generates a sorted Y matrix depending on a column of X matrix.
- With the method POD, a bad interpolation issue appeared when a reduced number of modes was used with “do training step” at “no” (save_file = -2).
ODYSSEE Nova – Optimizer
- Bug Fixes
- The final optimal point displayed in the summary through the command prompt was not the good one selected among the convergence points when we worked with multiple objective functions Fi. The log file has been updated to correctly display the optimal point, with constraints if there are.
- For an optimization problem with an equality constraint, the optimal point proposed at the end of the log file was not appropriate if the problem had 2 or more parameters.
- For the methods downhill, simulated, SQP, SLSQP and NLPQLP, after the convergence, the result
F (evaluation function) was missing in the summary of the optimal point.