Washout Identification and Repair

Purpose

A density or neutron log is only as good as the contact between the pad and the wall of the hole. Where the hole is washed out, rugose or caked, the readings are wrong, usually in a consistent direction (low density, high neutron), and they are wrong in exactly the intervals that are easy to mistake for real rock: shales and coals. This step finds those intervals, protects the real rocks that look like them (coal and salt), and decides what to do with the damaged data: remove it, or replace it by a prediction that is labelled as one.

Position in the workflow

Upstream. The flags need curves with standard names, units and depths (Curve Aliasing and Mnemonics, Log Curve Cleanup), including the caliper in inches and the bit size for each section.

Downstream. The result feeds Curve Normalization, where a washout in the reference interval would move the picks, and the whole of the workflow that uses density and neutron: porosity (Porosity), clay volume, mineralogy and, in source rocks, TOC from density (TOC Analysis).

Error propagation. A density that reads low by 0.1 g/cm³ gives a density porosity high by roughly 6 porosity units in a sandstone (0.1 divided by a density contrast of about 1.65 g/cm³), and the same low density is read as organic matter by a density-based TOC method. The error is not random: it is biased in one direction and concentrated where the hole is worst. A repaired curve passes its own uncertainty on to every later step.

Key concepts

Bad hole. A Bad hole is a borehole condition that damages the contact between a pad tool and the formation. The density, PE and neutron are most affected, the sonic and resistivity less, and the gamma ray little.

Two kinds of evidence. The caliper is mechanical evidence about the hole (Caliper-Based Flags). The logs themselves give evidence about the measurement: the density correction, range limits and consistency tests (Log-Quality Flags). Neither is conclusive alone.

Real rocks that look like a bad hole. Coal and salt have low density and odd neutron responses. They must be identified and exempted before the bad-hole flags are used (Coal and Salt Identification).

Padding. A flag marks the worst part of the damage. The tool starts to lose contact before the flag and recovers after it, so a flag is padded by a few samples on each side.

Remove or repair. Removal (Repair: Null-Out) leaves honest gaps. Repair replaces the gap with a prediction from other curves, by regression (Repair: Regression-Based) or by a machine-learning model (Repair: Machine-Learning Infill).

Prediction is not measurement. A repaired value is an estimate computed from other curves in the same well. It contains no new information about the rock, it can be wrong while looking right, and it must stay labelled.

Method selection guide

Method	Inputs	Use when	Strengths	Weaknesses
Caliper excess over bit size	Caliper, bit size per section	A good caliper is available and bit sizes are known	Direct and easy to explain	Misses rough but near-gauge holes and good-looking caliper with poor contact
Caliper rugosity	Caliper, a window	Breakouts, spiralling, rough walls near gauge	Finds what the excess test misses	Flags the edges of smooth washouts and not their middle; depends on window
Density correction, range tests	DRHO, RHOB, PE, NPHI, DT	No caliper, a poor caliper, or as a cross-check	Measures the log itself, curve by curve	Thresholds are tool-dependent; range tests flag real extremes
Coal and salt identification	Density, neutron, sonic, with PE and GR	Coal or salt is expected in the section	Stops real rock being nulled; gives a lithology mask	Overlaps with washouts; needs thresholds per tool
Null-out	Flag, padding	Default for any flagged interval	Honest: no invented data	Leaves gaps for later steps
Regression repair	Good curves in the same well, a training window	A continuous curve is needed over modest gaps in similar rock	Simple, transparent, gives a fit quality	Fails if predictors are damaged or the rock changes; carries no new information
Machine-learning infill	As above, plus a validation scheme	Continuous curves for display or a model that cannot take nulls	Handles non-linear relations	Looks plausible when wrong; smooths away thin beds; easy to validate dishonestly

Decision guidance

Flag with more than one kind of evidence. Use the caliper and the density correction together, and exempt coal and salt first.
Null by default. Repair only where a continuous curve is needed, and prefer the regression to a machine-learning model unless a non-linear relation is demonstrated.
Repair short gaps in uniform rock, and distrust a repair longer than a few tens of feet, across a lithology change, or in the reservoir.
If the bad hole is in the reservoir interval and matters for the answer, say so in the result. A re-log, another tool or an offset well is a better answer than any repair.

Shared parameter picking

Bit size. One value per hole section, in inches. It is the reference for the caliper tests, and a wrong bit size makes every flag wrong.

Limits. The caliper excess (Caliper excess limit) and rugosity (Caliper rugosity limit) limits and the density-correction limit are the main tuning parameters. Set them with a well that is known to be bad in places and good in others, and use the same values across a field unless the tools change.

Padding length. The number of feet by which a flag is grown on each side. One padding length is used for all curves flagged by the same hole, and for the nulling and the taper at the edge of any repair.

Order of operations. (1) Identify coal and salt; (2) flag bad hole outside them; (3) pad; (4) null; (5) repair if needed, trained only on good, unpadded samples. The order matters: padding before the coal and salt exemption can remove the edges of real beds, and repairing before padding trains on damaged samples.

Per-curve flags. Keep a flag per curve (density and PE, neutron, sonic). A bad density does not make the sonic bad.

Recommended default approach

Absent other information, a careful generalist would:

Check the bit size and the caliper in each section, and fix them before anything else.
Flag coal and salt where they are expected, with a vote among density, neutron and sonic, and PE as a tie-breaker.
Flag bad hole for density and PE with the caliper excess and the density correction together, and add rugosity if the hole is rough.
Pad the flag by about 2 ft on each side and look at several flagged intervals before accepting the padding.
Null the damaged curves in the padded intervals and keep the raw curves.
Repair only the intervals where a continuous curve is needed, by regression on undamaged curves, with a flag curve and a stated band.
Record the limits, padding, the share of each well flagged and the share repaired.

Combining methods

Flags from different tests are combined by OR: a sample is bad if any enabled test says so. Coal and salt flags override the bad-hole flag in the opposite direction, so that real coal and salt are not removed. Repairs are not combined: choose one method for an interval and record it. If two repairs disagree by more than their stated uncertainty, treat the interval as unrepaired. Averaging a regression and a machine-learning prediction does not make a measurement, and the two share their predictors, so they are not independent.

QC of results

A good result:

has flags that coincide with caliper excursions and large density corrections, and none in clean, in-gauge rock,
leaves coal and salt unflagged as bad hole and labelled in their own mask,
has nulled intervals that cover the whole damaged zone, including the edges,
shows no low-density tail in the cleaned density histogram of a shale,
has repaired intervals that join the measured curve smoothly, stay inside the range of good data, and are marked by a flag curve, and
reports the share of each well flagged and repaired.

Signs of a bad result: nulled coal beds, flags where the caliper is in gauge and the density is stable, damaged shoulders left on the edges of flagged zones, and repaired curves that are suspiciously smooth.

Common pitfalls

Using one bit size, or a fixed caliper limit, for a well drilled with several.
Trusting the caliper alone: a good caliper does not guarantee pad contact, and a bad one does not guarantee a bad log.
Nulling real coal, salt, pyrite or heavy-mineral zones as bad hole.
Flagging without padding, and training or averaging on the damaged edges.
Repairing with predictors that are damaged by the same hole (neutron, PE).
Training a repair on the whole well and not on a window near the gap, or across a lithology change.
Validating a machine-learning infill on a random split, so that the score reflects leakage and not skill.
Using repaired values as independent evidence, or without a flag, in a quantitative result.
Overwriting the raw curves.

Going Deeper

Borehole-environment correction and editing of density logs is as old as the pad tool. The modern practice of automated flags and repairs adds convenience and speed, and with them a risk: the more automatic the repair, the easier it is to forget that the repaired curve is not a measurement. Research on machine-learning infill of logs is active and the results are encouraging for display and for training other models, and are mixed where a quantitative answer rests on the infilled interval. A stable principle is to spend effort on prevention and detection (hole condition, tool choice, quality curves) and to keep any repair visible and its uncertainty stated.