Field Data, Drilling Logs and Lithological Classification — Artificial Intelligence in Geological Engineering

Gains:

Ability to structure and lithological classification of drilling logs, core descriptions and field notes with AI
Ability to convert free text field observations into standard geological terminology and tables
Ability to verify AI-generated lithological interpretations with core photography, log curves and stratigraphic plausibility

The raw material of geological engineering is data, and most of this data is born irregularly. Core descriptions kept at a site are free texts written in different words by different geologists; The logs of a drilling campaign were kept in different formats, in different units, sometimes manually and sometimes digitally; The notes in the field notebook are full of abbreviations and personal habits. In this unit, we cover the most concrete and safest area of profit of artificial intelligence: structuring scattered field data (translating it into queryable tables), standardizing it (bringing it to common terminology) and making preliminary lithological classification (rock type separation). But let's say it from the beginning: artificial intelligence organizes the data, the engineer approves the geological interpretation.

Let's clarify two key terms first. A borehole log is a record of properties measured against depth along a well; these measurements can be geological (lithology, weathering) or geophysical (gamma ray, resistivity, density). Core (English core) is a cylindrical rock sample extracted from underground by drilling; It is the most valuable evidence as it allows direct observation of the underground. Lithology is the type and physical character of the rock (such as sandstone, limestone, shale).

Structuring Free Text

A field geologist writes a description: "0-3.2 m brown, loose, clayey gravel, poorly sorted; 3.2-7.5 m gray, moderately weathered, fine-medium grained sandstone with sparse carbonate cement; 7.5-12 m light gray, solid, fissured limestone, calcite vein at 9 m." This single sentence is actually a compressed version of a table: depth range, color, degree of weathering, lithology, grain size, secondary properties. Artificial intelligence decompresses this.

The critical point is this: ask the model for transformation, not inference. In other words, place the information clearly written in the text into the fields; Do not make up what is not in the text. For the "degree of weathering" field, if the geologist wrote "medium weathered", the model writes this; If the geologist did not specify any, the model should write "unspecified" and not fit a value. This distinction is the basis of data integrity.

Step by step: turning a core book into a spreadsheet

You define the schema first. Which columns will be: depth_top, depth_bottom, lithology, color, grain_size, weathering, discontinuity, secondary_mineral, note. You impose the schema on the model so that the output is consistent.
Give the standard dictionary. Give the allowed terms for lithology (sandstone, siltstone, limestone, marl, claystone, conglomerate...) and the scale for weathering (fresh, light, medium, high, full). Thus, expressions such as "a little weathered" become standard.
Process small batch. Give the entire notebook not at once, but in batches of 20-30 lines; You catch mistakes early.
Go back to the source and verify. Compare random rows of the resulting table with the raw text; especially depth limits and lithology terms.

Lithological Classification: From Logs to Interpretation

AI supports lithology in two ways. The first is to standardize the definition of geologist as above. The second and more advanced is lithology estimation from geophysical log curves. For example, the gamma ray log (the natural radioactivity of the rock) increases with clay content; Clayey levels give high values, clean sandstones and limestones give low values. By adding resistivity and density logs, the model can suggest lithology boundaries from curve patterns.

The limit here is important: logs are indirect measurements. High gamma can be clay, feldspar-rich sand or volcanic ash. Therefore, log-based lithology prediction is a hypothesis; It is not definitive unless it is connected with core or cuttings (rock chips that come to the surface during drilling). When the model draws a boundary, ask if there is any core at that depth; If there is, the core wins.

Three Mini Cases: By the Numbers

Case 1 — Standardization consistency. In a mining company, 6 different geologists used 11 different expressions for "weathered granite" (weathered, altered, rotten, clayey...). Artificial intelligence-assisted standardization placed the 4,800-row database on a 5-level decomposition scale; Afterwards, the query "medium-highly segregated zones" took seconds, whereas previously it took days. A random 10% of the standardization was checked by core, the agreement was high.

Case 2 — Log correlation preliminary draft. Lithology correlation would be made from gamma logs in two neighboring wells (150 m apart). The model proposed a preliminary correlation that matched four common levels and flagged a thickness difference at one level as "possible fault or wedge-out." When the engineer looked at the core of the marked zone, he found a really small fault; The model didn't give the definitive answer, but it drew attention to the right place.

Case 3 — Capturing the fabricated border. In one well, the model assigned a “basalt” lithology to the 40-45 m interval for which there was no core data. Basalt was not expected in the region; there was only gamma increase in the recording. The geological plausibility filter was activated, a clast sample was requested from that interval, and the level was found to be clayey siltstone. The model had misinterpreted the high gamma.

Weak Prompt / Strong Prompt

Weak prompt:

Convert these core definitions into a table.[text]

Powerful prompt:

The powerful prompt imposes the schema, the dictionary, and the "fabrication ban" together; The output is both consistent and auditable.

Four Copiable Templates

1) Field notebook standardization:

Translate the following field note into standard terminology. Open abbreviations(e.g. "stone"->"sandstone"), but DO NOT change the content. Mark ambiguous phrases as "[ambiguous]" and retain the original phrase in parentheses. Note: [field note]

2) Log-based lithology hypothesis:

Suggest possible lithology boundaries from the following log data (depth, gamma, resistivity, density). For each boundary, write: (a) from which curve change you derive it, (b) the possibility of an alternative lithology, (c) the need for coring verification. Do not present any boundary as "certain." Data: [log table]

3) Two well correlation sketch:

Compare Well-A and Well-B logs. List matching levels, matching reason, and mismatched/suspicious zones separately. Label fault, erosion or facies change as "must be verified". Data: [two well logs]

4) Data quality control:

Look for logical errors in the lithology table below: overlapping depth ranges, gaps, impossible alignments, unit inconsistencies, unauthorized terms. Report findings as a "row number | problem | suggestion" table. DO NOT CORRECT data, just mark it.Table: [lithology table]

Authentication Proof Hierarchy

Evidence type

directness

reliability

When to use

Core (core)

most direct

highest

Critical lithology boundaries

Crumb (cutting)

medium

high

In the absence of core

geophysical log

indirect

medium

Continuous profile, gap filling

Model prediction

derived

Low (alone)

Hypothesis, screening

Free text note

Variable

Depends on geologist

Additional context

Tip: When helping artificial intelligence find lithology boundaries, ask for "possible boundary and confidence level", not "exact boundary". The core gives accuracy, not the model. The output then becomes a screening tool, not a false claim to certainty.

Caution: The model tends to fill in depth ranges that are not in the data to "leave no gaps". It can produce fake lithology, especially in zones with core loss. Clearly mark core loss zones and instruct the model to “leave blank if no data”.

Common mistakes

Seeking inference rather than transformation. Expecting the model to "guess" what is not in the text; This opens the door to fake data. Just structure what is written.
Not providing a standard dictionary. Allowing free terminology; The result is an unquestionable, inconsistent picture.
Accepting the log estimate as accurate. Considering indirect measurements (gamma, resistivity) alone as evidence of lithology; To move the borders to the signature without binding them with core/crumbs.
To hide core loss zones. Filling data gaps in the model; space must remain space.
Processing in batches and never checking. Not translating thousands of rows at once and doing random sample checking.

In summary

Artificial intelligence provides safe and high efficiency in structuring, standardizing and preliminary lithological classification of dispersed field data.
Ask for transformation from the model, not inference; Do not make up what is not in the text, print "unspecified".
The log-based lithology estimate is indirect and a hypothesis; It cannot be confirmed until it is connected to core and crumb samples.
You impose the schema and standard dictionary; Only then will the output be consistent and controllable.
Core drilling is at the top of the evidence hierarchy; At the bottom, the model prediction alone is insufficient.

Application task

Take 15-20 lines from a field notebook or core description you have (anonymize confidential field information). First define your own schema and allowed lithology/weathering dictionary, then translate it into a table with the powerful prompt. Compare each row of the resulting table to the raw text and flag any areas that the model has fudged or misplaced; Check if at least one "unspecified" field is left blank correctly.

checklist

[ ] I can convert free text field observations into a table structured with diagrams and standard dictionary.
[ ] I apply the "unspecified" rule that I want a transformation from the model and that I don't make it fit something that doesn't exist.
[ ] I understand that the log-based lithology estimate is an indirect hypothesis and should be linked to the core.
[ ] I mark core loss and data gap zones and prevent fitting.
[ ] I validate the output according to the evidence hierarchy (core > crumb > log > model).