SimPEG_Scripts

Repository of scripts to run SimPEG inversions and forwards

To run an inversion, open a command terminal and type:

python PF_inversion.py your_project_directory\PF_inversion_input.json

Parameters

• data_type: str

  Data file format,

  • ubc_grav: See GRAV3D documentation
  • ubc_mag: See MAG3D documentation

• inversion_type: str

  Inversion type

  • 'grav': Invert for density in (g/cc).
  • 'mag': Invert for magnetic susceptibility in (SI).
  • 'mvi': Invert for effective susceptibility in Cartesian coordinates.
  • 'mvis': Invert for effective susceptibility in Spherical coordinates.

• core_cell_size: list

  Dimensions of the smallest cell size in the mesh. [dx, dy, dz]

Optional settings: type = DEFAULT

• add_data_padding: bool = False

  Add an area of data padding around the input survey to manage edge effects

  Currently only enabled for Geosoft grid imports

• alphas: list = [1, 1, 1, 1]

  Alpha weights used to scale the regularization functions.

  • Scalar (density, susceptibility): Requires 4 values for [a_s, a_x, a_y, a_z]
  • Vector (mvi): Requires 12 values for [a_s, a_x, a_y, a_z, t_s, t_x, t_y, t_z, p_s, p_x, p_y, p_z]

• alphas_mvis: list = [1, 1, 1, 1, 0, 1, 1, 1, 0, 1, 1, 1]

  Alpha weights used to scale the regularization functions only used for mvis inversions.

  • Vector (mvi): Requires 12 values for [a_s, a_x, a_y, a_z, t_s, t_x, t_y, t_z, p_s, p_x, p_y, p_z]

• chunk_by_rows: bool = False

  Alternate memory management mode that can be faster for very large sensitivity or forward calculations

• depth_core: dict: {str, float} = {"method": value}

  Thickness of core region defined by method:

  • value: Set value in meters
  • auto: Compute mesh depth as: value * survey width.

• decimate_to_mesh: bool = False

  Downsample the input data to at most 1 observation above each grid cell. This significantly reduces the problem size when using padding, as well as when core_cell_size is larger than the data spacing.

  Currently only enabled for OcTree meshes

• detrend: dict {str: int} = None

  Remove trend from the data {method: order}.

  method

  all: All points used corners: Points on the convex hull

  order

  Integer defining the order of polynomial: 0, 1 or 2

• drape_data: float = None

  Value defining the drape height above topography. If drape_data is used, the Z elevation of receivers are over-written

• input_mesh_file: str = None

  Mesh file in UBC format used to load the model_reference and model_start. The same mesh is used for the inversion if inversion_mesh_type is same type.

• inversion_mesh_type: str = "tree"

  Type of mesh to be used in the inversion. If type differs from the input_mesh_file, then the input model_reference and model_start are transfered over using a NearestNeighbour interpolation.

• inversion_style: str

  Inversion style chosen from:

  • voxel: Standard voxel base inversion [DEFAULT]
  • homogeneous_units: Invert for best-fitting value for each domain defined by the unique values found in model_start.

• lower_bound: float = -Inf

  Value to use for lower bound in each cell

• max_chunk_size: float = 128

  Size of data chunks to store in memory

• max_RAM: float = 4

  Maximum available RAM. If tiled_inversion is True, then the tile size will be defined to keep the problem smaller than max_RAM

• model_norms: list = [2, 2, 2, 2]

  Model norms to apply, in range of 0-2 where 2 is least-squares

  • Scalar (density, susceptibility): Requires 4 values for [Lp_s, Lp_x, Lp_y, Lp_z]
  • Vector (mvi): Requires 12 values for [Lp_s, Lp_x, Lp_y, Lp_z, t_s, t_x, t_y, t_z, p_s, p_x, p_y, p_z]

• model_start: str or float or list[float] = 1e-4

  Starting model to be loaded with the input_mesh

  • str = filename: Load starting model from file. If inversion_mesh_type differs from the input_mesh_file, the model values are interpolated to the new mesh.
  • float = value: Start property, scalar [1e-4].

  • list = [value, value, value]: Start property model values for vector model [1e-4, 1e-4, 1e-4]

    If scalar input used for vector model, assume scalar amplitude in inducing field direction.

• model_reference: str or float or list[float] = 0

  Reference model to be loaded with the input_mesh

  • str = filename: Load starting model from file. If inversion_mesh_type differs from the input_mesh_file, the model values are interpolated to the new mesh.
  • float = value: Start property, scalar [1e-4].

  • list = [value, value, value]: Start property model values for vector model [1e-4, 1e-4, 1e-4]

    If scalar input used for vector model, assume scalar amplitude in inducing field direction.

• input_vector_type: str = 'xyz'

  Specify vector file format used for model_start and model_reference (if it has a .FLD extension)

  • str = 'xyz' if vector files are XYZ (same as vector model output format). 'atp' if vector files are AMP, AZI, DIP.

• new_uncert: list = [0, 1]

  List of values to be used for uncertainties set as [%, floor] (% from 0-1) where uncertainty = max(% * |data|, floor)

• no_data_value: float = -100

  Value to use for no-data-value

• parallelized: bool = True,

  Use dask parallelization

• result_folder: str = "SimPEG_PFInversion"

  Directory used to output the results

• show_graphics: bool = False

  Show graphic plots

• target_chi: float = 1

  Target chi factor

• tiled_inversion: bool = True,

  Use tiles to speed up the inversion and keep the problem small

• tiling_method: str = 'cluster',

  Choice of methods to brake up the survey into tiles:

  • orthogonal: Splits the survey by adding tiles of equal spatial dimensions.
  • clustering: Use scikit-learn.AgglomerativeClustering algorithm (slow for large problems)

• upper_bound: float = Inf

  Value to use for upper bound in each cell

Magnetic only

• "inducing_field_aid": [TOTAL FIELD, DIP, AZIMUTH], New inducing field as floats

More documentation to come in 2020!! Stay tunes.

Notes on Octree Mesh

The following parameters can be used to modify the octree mesh.

• core_cell_size : Smallest cell size dimension (h_x, h_y, h_z)
• octree_levels_topo : Number of cells inserted below topography [*]
• octree_levels_obs : Number of cells inserted below the data points [*]
• octree_levels_padding : Number of padding cells inserted horizontally around the data points [*]
• max_distance : Maximum triangulation distance used by the refinement
• depth_core : Minimum depth of the mesh below the lowest point
• padding_distance : Minimum padding distance along the cartesian axes

[*] List of integers [nC_1, nC_2, ... ] ordered from the lowest octree level (smallest cell size) to the highest. For instance the list [2, 6, 10] will request for at least 2 fine cells (h_x), followed by 6 cells at the 2^{th} level (2^1*h_x) followed by followed by 10 cells at the 3^{th} level (2^2*h_x).

See the refine_tree_xyz documentation for more details and examples.