What's New in ODYSSEE CAE 2022.3

ODYSSEE Lunar – GUI

• MLP solver is now available among Lunar’s prediction methods. The solvers configuration dialog box has been updated accordingly with the following options:
  • Choice of Activation function between linear or sigmoid.
  • Choice of Optimizer between Adam and SGD.
  • Choice of MLP Method between Standard (no constraints on Y ̂) and MINN (Mechanics Informed Neural Network) (with constraints).
  • Choice of Normalization between: None, Normalize, Standardize or Center.
  • Add one or several layers, and for each layer, to adjust the number of neurons.
  • Define the number of iterations.
  • Adjust the percentage of validation cases used to train the neural network.
  • Other settings.
• In the LS-Dyna parser tool, it is now possible to extract outputs from <matsum> and <dbfsi> files.
• In the Adams Car parser tool, the copy of the Car Database folder contained in each subdirectory M1, M2, M3… is now removed automatically at the end of the parsing process. This saves a large amount of storage space.
• Updated the default parameters of SVM solver: Type: Regression, Kernel= Polynomial, S1=1, S2=1, and Gamma= 0.5.
• Animation:
  • Display the elemental results from a D3plot for only one part of the model.
  • Display the elemental results for a reduced D3plot. The min/max range in Lunar is based on the values found for the visible elements and also the deleted elements.
• User Script:
  • The user script UserScript_BestMethod_by_Cross_Validation_on_X.qsr has been updated:
    • To allow using the Clustering method on the scalar values of Y data.
    • When the criteria error norm L2 was selected, the worst result was displayed instead the best.
    • The fourth argument has been edited for clarity: Previously you had to enter a value defining the cross validation i.e., the number of cases desired per blocks. Now you have to just enter the required number of blocks.
  • Updated the UserScript_LeaveOneOut.qsr user script.
    • Added a personal predictive method with an id 15. You can update this one to define your own method.
  • New User script UserScript_Split_Database.qsr user script.
    • Split the dataset into training and validation sets and generate the corresponding X, Y, Xn and YN files automatically.
  • Updated the Lunar Help document.

• Bug Fixes
  • FMU:
    • A line was added at the end of the X file in the exported FMU v3.0, leading to a license error message.
    • The model FMU v2.0 for an ADAMS model didn’t work when it is from the version 2022.2.
    • For different XN values, FMU v2.0 gave the same results for some methods.
  • The MLP solver must automatically change the save files at -1.
  • Fixed text in the stationarity list of Kriging.
  • Animations prediction with POD method and “recompute database files” option unticked was slower than with “recompute database files” option ticked.

ODYSSEE Quasar – Solver

• Add a new method for Multiple Layer Perceptron (MLP) named Mechanics Informed Neural Network (MINN) applying a constraint on the Jacobian of the output layer .
• FMU improvement: Use the analytical form of the stiffness matrix when using MINN. For the case of Standard Neural Network (MLP), we use finite difference method to compute this matrix.
• A message is now displayed in the command prompt if the user calls a bad *.qsr file (error in the filename or file not existing in the directory).
• Update/New Quasar functions:
  • For DOE:
    • Add automatically new points in a sample file (X.csv). For simplify the GpNewpoint function, the arguments list has been reduced. GPNewpoint becomes GpAdaptative: ext("operator","matrix","gpadaptive",X,Xcandidate,gridPts,n_pts,PCA)
      The new function gives the same results of the previous one.
    • New keyword to generate a “Normal” distribution with a specific mean and variance. To control the space filling, you need to enter Kscale=4 and load csv file File_MeanSTDV containing 2 columns: first the means, second the standard deviation.

ext("doe","meanstddevnormal",Nvariables,Nsites,Neighbours,Seed,Kscale=4,"doe_mstdnormal_sampling",File_MINMAX,rmin,rmax,File_MeanSTDV)

• For regression:
  • ext("mining","effectsqr",Xstd,Y,"regressionQR_"). Add an output file (regressionQR__REGRESSION_COEFFICIENTS.csv) for the effectsqr function.
  • When ARBF interpolation is applied, an output file named out is generated saving the chosen base function and radius.
  • GPNewpoints external function to add samples automatically to the database becomes GPAdaptative external function simplifying the arguments list.

ext("operator", "matrix", "gpadaptive", X, Xcandidate, gridPts, nbr_new_points, nbr_Eigen_vector_PCA)

• A new neural networks function called NN has been implemented for new options integration.

Matrix YN = NN(X, Y, XN, architecture, learning_rate, activation, optimizer, percentage, isshuffled, n_epochs, batch_size, Prefix, "normalize ", method, save_files)

• The arguments of LSTM function are reordered for a best intuitiveness.

Matrix YN = LSTM(X, Y , XN, n_neurons, Tx, batch_size, learning_rate, activation_name, optimizer, normOption, percentage, isShuffled, n_epochs, prefix_file, saveFile)

• For pre-processing:
  • Added the matspline(X,Y,XN,"root.csv") function that will transform a curve or a signal with n points in a curve or signal of m points where m > n.
  • Added a flag argument ext("operator","matrix","order_y_by_row_index_X", X, Y,column=2, flag). If FLAG=0 all columns of Y will be ordered as in X (:,column) else FLAG=1 only the column col of Y will be reordered.
• Added 3 new functions in the Useful_Functions.qsr library:
  • Replace function to change a text from start position to an end position.
  • Order function to sort by column a Matrix at 1 line (equivalent to a vector).
  • addLabelToMatrixFileAndSave function to save a matrix in a file with the asked separator and columns labels wanted.
• Bug Fixes
  • The multiple successive calls of the SVM function generated an error.
  • For the correlation and covariance functions, if matA, and matB tested have not the same size, then an error message is now displayed.

ODYSSEE Nova – Optimizer

• Bug Fixes
  • In the optimization log file generated by Nova, the proposed optimal point is the same of the Lunar's or near at one or two iterations.
  • The optimization results were not displayed in the console at the end of the optimization when using Downhill Simplex algorithm.