Pytorch Nested Cylinder Models

Straightfoward CNNs were developed in pytorch to estimate the scaling of a Preston-Tonks-Wallace (PTW) strength model from a nested cylinder experiment.

Model Inputs & Outputs

Networks trained on nested cylinder data take as input an image showing the density throughout the experiment. An example density field is shown below.

These models predict the value that the PTW strength model was scaled by.

Running the Nested Cylinder Example

To run the scripts on the pytorch nested cylinder networks, use the following arguments:

• --PACKAGE pytorch

• --EXPERIMENT nestedcylinder

• --MODEL ../examples/pyt_nestedcyl/trained_rho2PTW_model.pth

• --INPUT_FIELD rho

• --INPUT_NPZ use any .npz file in ../examples/pyt_nestedcyl/data/

• --INPUT_DIR ../examples/pyt_nestedcyl/data/

• --DESING_FILE ../examples/pyt_nestedcyl/nestedcyl_design_file.csv

• --FIXED_KEY id0643, --FIXED_KEY idx00130, or --FIXED_KEY None

Nested Cylinder Data

The file names of nested cylinder data contain multiple pieces of information about their contents. The two relevent componets are the id and the idx:

• The id is followed by a number identifying the experiment ID.

  • The value of PTW_scale is identical across files with identical id ‘s.

  • Data is included for id0643 at all 22 timesteps.

• The idx specifies what time step the simulation was at.

  • The value of the sim_time will be identical across files with identical idx ‘s.

  • Data is included at idx00112 for 31 differnt simulations.

Each .npz nested cylinder data file contains the following fields:

• sim_time (scalar): simulation time stamp corresponding to a unique idx value

• rho (2D array): density training field (use -IN_FIELD rho)

• hr_outerWall, hr_bottomWall, hr_mcSide, hr_mcBottom, hr_innerCylClide, hr_innerCylBottom, hr_MOI, hr_innerCylBottom, hr_innerCylCorner, hr_innerCylRight (2D arrays): density of simulation compoments

• volhr_outerWall, volhr_bottomWall, volhr_mcSide, volhr_mcBottom, volhr_innerCylSide, volhr_innerCylBottom, volhr_MOI (2D array): volume average density of similation components

• pressure (2D array)

• temperature (2D array)

• melt_state (2D array): binary array of if a cell has melted

• porosity (2D array)

• eqps (2D array): equivalent plastic stress

• eqps_rate (2D array): equivalent plastic stress rate

• eff_stress (2D array): effective stress

• bulk_mod (2D array): bulk modulus of the material

• sound_speed (2D array): speed of sound in the material

• rVel (2D array): veliocty of material in the R-axis direction

• zVel (2D array): velocity of material in the Z-axis direction

• Rcoord (1D vector): vector containing position in cm of all cells along the R-axis

• Zcoord (1D vector): vector containing position in cm of all cells along the Z-axis

Model Architecture

These models consist of a single branch that passes the image input straight forward through the network.

Model Layers

The layers in model follow the following naming convention:

• in???: layer near top of the network

• ??Conv??: 2D convolutional layer

• ??Norm??: 2D batch normalization layer

• ??Activation: GELU activation layer

• interp_module.??.##: layer in “interpretability stack”

• reduction_module.??.##: layer in the “reduction stack”, which reduced layer size by using a stride ≠ (1,1)

• end???: layer near the end of the model

• hidden: linear hidden layer

• hidden#Activation: GELU activation layer after a hidden layer

• linOut: linear layer that generates output

• flattenLayer_##: torch.nn.Flatten() layer