Performance improvements using FIELDS. Suggestions

[oddvarlia]

Drogon and the RMS plugin for APS now utilize an ERTBOX grid for FIELDS. This ERTBOX grid is created in RMS. It is a 'shoebox' grid with NX, NY equal to the modelling or simulation grid while NZ for this grid is user defined and must be at least as large as the maximum possible number of layers in the zone in the modelling grid with most grid layers. So if the structural uncertainty creates variable number of grid layers, the ERTBOX grids number of layers must be maximum number of layers any structural realization may give. The geomodelling/simulation grid may be created with different types of conformity for each zone:

• Proportional gridding (Number of layers are fixed for the zone)
• Base conform gridding (Layer thickness is constant and number of layers for a given zone may vary from realisation to realisation. The grid zone is built from base of the zone and upwards)
• Top conform gridding (Layer thickness is constant and number of layers for a given zone may vary from realisation to realisation. The grid zone is built from top and downwards.
• • Mixed conform gridding. The grid layers have constant thickness. The layering is squeezed between two surfaces that are not the same as the zone top and zone base surfaces.

The main purpose of using the ERTBOX grid instead of the geomodelling grid/ simulation grid is to avoid the practical problems with sampling effects and large grids for fields covering the whole modelling grid instead of individual zones.

The current workflow is as follows when using RMS APS GUI plugin together with FMU/ERT:

• Create the ERTBOX grid and rotate and place it such that it is as far as possible rotated in the same ways as the modelling grid and in map view located where the modelling grid is located. The NX, NY are determined by the geomodelling grid. If the modelling grid contains reverse staircase faults, the NX and NY may be larger than the 3D modelling grid since the simbox grid size is used. (In RMS the simbox size and the read grid size is found under Information menu related to the grid).
• When creating initial ensemble: APS simulates GRF's (Gaussian Random Fields) for the ERTBOX grid, export it in GRDECL or ROFF format so that ERT FIELD command can pick it up.
• Depending on the grid conformity of the modelling grid, the ERTBOX grid is filled from top or from bottom. All grid cells in ERTBOX grid that corresponds to inactive grid cells or collapsed grid cells of the modelling grid will anyway get som values for the GRF's since all grid cells in the ERTBOX grid will be simulated. Even with parametrized trends for GRF (mathematically defined trend functions) the trend value can be calculated for all grid cells in the ERTBOX grid.

About the performance issue:

1. Since the ERTBOX grid is used, only the GRF's relevant for each modelling grid zone is created. This is an advantage. If e.g. the modelling grid as 3 zones, A, B,C but only zone A and B are modelled with APS, it is only necessary to create the N GRF fields for A and the M GRF fields for B, in total N+M GRF fields, all with size NX x NY x NZ where NZ is the ERTBOX number of layers. It is not necessary to have GRF values covering zone C or equal number of GRF fields for all zones.
2. The disadvantage using one ERTBOX size for all zones is that NZ may be much larger than necessary for thin zones and therefore allocate space for many GRF cell values that are unphysical (not belonging to any physical grid cell in the geomodel/simulation grid. In the example above the number of model parameters introduced for ERT using FIELD will be (N+M) x NX x NY x NZ. If the maximum number of grid layers necessary for zone A is 5, and for B is 50, the number of values using the current solution with ERTBOX will be (M+N) x NX x NY x 50 while the necessary maximum could have been N x NX x NY x 5 +M x NX x NY x 50
3. A solution to this may be:
   • Introduce not one common ERTBOX that is re-used for all zones, but a stack of one ERTBOX grid per zone with the ERTBOX size NX x NY x NZ_zone_specific. In practice this can be done by introducing one ERTBOX grid with multiple zones stacked on top of each other. This is very analogous to what RMS can visualize when visualizing a grid in simbox view. The most important difference is that the number of layers for each simbox (each zone) may be larger than what RMS uses in its simbox for the zone since the ERTBOX zone must be able to save enough space for any realisation of the structural model.
   • Introduce a table defining which zones have which GRF field and how many layers the zone needs.
   • Modify the FIELD keyword in ERT to understand that the specified field input is only for a specified zone, not the whole stack of ERTBOX grids.

This proposal may need to be discussed to see if it is can be prioritized or not and how much work it will involve modifying the ERT code for the FIELD command and the consequence of the internal representation of the field data in ERT.

[dafeda]

ERTBOX is not a concept defined in ERT's code as far as I can tell. Do you have some more information on this @oddvarlia ?

[oyvindeide]

Thanks for the suggestion! I think having one GRID per FIELD is the most actionable at the moment, so I split that into a separate issue: #8418, and will close this issue.