Physics class + Lassen compatibility

[dreamer2368]

Lassen compatibility

Some prerequisites are down-graded to be compatible with Lassen OpenCE-1.9.1. The burgers 1d equation example is working on LC Lassen. Installation procedure is added in README.

Physics class

FOM simulation solver is now modularized into Physics class. Currently only Burgers1d class is provided as a derived class.

Main member variables:

• dim: physical space dimension
• qdim: solution (denoted as q) dimension
• grid_size: the shape of the grid nd-array.
• qgrid_size: the shape of the solution nd-array.
• x_grid: spatial coordinates in numpy nd-array, assuming the shape of:
  • 1d: (grid_size[0], )
  • 2d: (2, grid_size[0], grid_size[1])
  • 3d: (3, grid_size[0], grid_size[1], grid_size[2])
  • higher dimension...
• nt: number of time steps
• dt: time step size
• t_grid: time grid in numpy 1d array
• offline: boolean to indicate whether the class is for offline simulation

Main member functions:

• initial_condition(self, param): returns the initial condition according to the parameter param
• solve(self, param): returns the time history of solution according to the parameter param
• export(self): returns a dict that stores the class information.
• generate_solutions(self, params): generates a list of solution time histories, given the list of parameters params

Refactoring physics from other classes

• BayesianGLaSDI is now initialized with the Physics class and owns it. All physics-relevant information is delegated to the Physics class.
• Autoencoder class now does not take FOM dimension as input, rather queries from the Physics class.
  • Autoencoder.encoder and Autoencoder.decoder now respects the shape of the FOM solution shape qgrid_size, in their input/output.

InputParser class

InputParser parses a dict, and provides a convenient interface to search multi-level keys, with a capability of default value and data type enforcement.

Activation function input for Autoencoder

At its initialization, autoencoder now looks for the option for activation function. Currently sigmoid and softplus are available. By default it take sigmoid.

[chldkdtn]

@dreamer2368 This looks great. Can you add in README file the instruction of running the new codes?

[chldkdtn]

@xiaolong7 and @TranApril if you can also keep track of this re-factoring of GPLaSDI code, that would be wonderful.

[dreamer2368]

@dreamer2368 This looks great. Can you add in README file the instruction of running the new codes?

README is now updated.

[dreamer2368]

@chldkdtn , Zoe was also able to run this branch on Lassen using opence 1.9.1 and visualize the results. Should we merge this branch now?

[chldkdtn]

@chldkdtn , Zoe was also able to run this branch on Lassen using opence 1.9.1 and visualize the results. Should we merge this branch now?

Okay, let's merge it now. @andersonw1 can catch up on this later.