Interpolation Troubleshooting
Note
Coming in Spatial Integration version 3.
Symptoms, what they mean and what to do about them.
Before You Start
"No drill datasets are available." Collars, surveys and a downhole table must be loaded first. Use the Drill Manager.
"This dataset has no numeric columns to interpolate." Only floating point columns can be gridded. If a numeric column was loaded as text, reload the table with the correct type - text columns are still usable as a Domain filter.
The sample count is far lower than the row count. Check the Nulls figure and your domain filter. Rows failing the domain filter are excluded silently by design, so a filter that matches nothing leaves you with nothing.
The Variogram Looks Wrong
"No lag bin reached 10 pairs." The lag bins are too narrow for the data. Increase the lag distance, add lags, or widen the cone tolerance if you are in directional mode.
Most points are hollow amber (fewer than 30 pairs). Same cause, less severe. Increase the lag distance so each bin gathers more pairs, or reduce the lag count so you are not reaching out past your data.
The curve is flat at the sample variance from the first lag. There is no measurable structure at the separations you are looking at. If this happens on a downhole variogram, the values really are noise-dominated at sample scale. If it happens only between holes, your holes are further apart than the range, and no interpolation will produce meaningful detail between them.
The curve is still climbing at the last lag. The range is longer than the distances you have measured. Add lags before fitting, or you will fit a range that is simply the end of your plot.
The curve rises, flattens, then rises again. Two scales of structure, or a regional trend. Try a tighter domain, or the Linear drift option on the Run tab.
The nugget comes out implausibly high. You are probably fitting a variogram whose shortest lag is a whole hole spacing. Switch Pair mode to Downhole only, set the lag distance near your downhole sample spacing, fit, and then tick Hold on the nugget before returning to the other modes.
"Computed from a deterministic subsample." Above 20,000 samples the variogram is computed from a repeatable random subsample, because every sample is paired with every other and the time grows rapidly. The result is representative; if you need the full set, use a domain filter or a depth range to reduce the sample count instead.
The anisotropy scan suggests a direction that contradicts the geology. Check the pair counts of the longest and shortest rows in the table. A single fence of closely spaced holes will produce a strong apparent anisotropy along the fence. Widen the cone tolerance, coarsen the azimuth step, or enter the orientation from the geology by hand.
The Run Fails or Refuses to Start
"Run is blocked: ..." The message names the cause. The usual ones are that samples are still being extracted, that the selection changed after they were extracted (re-extract), that a quality check is still running (fits run one at a time), that the output path is empty, or that a parameter is invalid.
"The kriging system is near-singular." Almost always a zero nugget with a smooth kernel - a Gaussian kernel with nugget 0 is the classic case - or two samples at exactly the same position. Set a nugget of about 1% of the sill and run again.
"The fit did not converge." Raise the Fit tolerance, raise Max iterations, and check for duplicate or contradictory samples at the same location.
"Ran out of memory." The grid or the sample set is too big for the machine. Coarsen the cell size - remember that halving it multiplies the node count by eight - or reduce the extent.
The fit phase seems to hang. No progress can be reported during the fit, so only the elapsed time moves. Watch the log pane, where the fit reports its progress. Cancelling during this phase can take a little while to take effect.
The Result Looks Wrong
Bullseyes around individual holes. The range is too short for the drill spacing, so each hole only influences its immediate surroundings and the rest drifts to the mean. Lengthen the range, or accept that the drilling does not support a continuous estimate at this spacing.
The grid is a smooth blur with no detail. Over-smoothing. Check the cross-validation slope and the variance ratio on the Quality tab. In order of effect: describe the anisotropy, reduce the nugget, shorten the range.
Grade smeared across a geological contact. The run mixed two populations. Use a Domain filter and interpolate each domain separately.
Values well outside the data range. Look for a Linear drift setting that is extrapolating a trend, or a Gaussian or high-numbered spheroidal kernel overshooting. Ordinary kriging with a spherical kernel is the conservative combination.
High grades diluted, low grades inflated - a low regression slope. Conditional bias from over-smoothing. As above, and worth testing with a stronger anisotropy ratio in cross validation before committing to a run.
Big blank areas after loading. The mask output is set to Blank nodes beyond threshold and much of the grid is beyond it. Either that is the truth about your data coverage, or the threshold is too tight - it defaults to one mean hole spacing. Switch to Write distance array to see the values and judge for yourself.
The grid mean is well above the declustered sample mean. A clustering or domaining problem more often than an estimation one. Check that preferentially drilled high-grade zones are not dominating, and consider a domain filter.
Performance
Everything takes too long. Work coarse while tuning. Node count drives the evaluation phase, and sample count drives the fit and the variogram - variogram time grows with the square of the sample count. A coarse grid and a domain-filtered sample set make the experimental cycle fast; refine only once the model is settled.
A quality check will not start. Fits run one at a time - a message says so. Wait for the run, cross validation or residual calculation in progress to finish.