Gains:
- Ability to clean drilling log, lithology and grade data with AI support and detect inconsistencies and coding errors
- Ability to accurately position AI as a pre-classification and hypothesis generator when interpreting core photography and geological observation
- Ability to cross-validate AI's inferred lithology/alteration interpretation with geologist observation and laboratory analysis
The basis of every mining project is the samples taken from underground. The most important of these examples is the drill core (a continuous sample that the drilling machine extracts from the rock in a cylindrical shape). The geologist examines the core; It records the rock type (lithology), structure, alteration (change of the rock by hot waters and chemical effects) and mineralization, and analyzes laboratory samples for grade. The result is a huge table recorded meter by meter, called a drilling log. There may be hundreds of drillings and tens of thousands of lines in a single project. Cleaning this data, combining it, catching inconsistency and extracting initial interpretations; This is an area where AI is really accelerating, but each step needs to be verified with a geologist's eye. In this unit you will learn how to clean drilling/geology data with AI, how to pre-interpret the core image and how to validate the results.
Anatomy of drilling data tables
A drilling data set typically consists of four tables, and combining them correctly is half the job:
- Collar: The starting coordinate (X, Y, Z) and total depth of each borehole.
- Survey (deviation): Change in slope (dip) and direction (azimuth) of the drilling along the depth. Drilling never goes perfectly straight.
- Lithology: Rock type and geological observations throughout the depth range (from-to).
- Assay: Laboratory grade results across the depth range (e.g. Au g/t, Cu %).
These tables must comply with rules such as from < to, ranges must not overlap, and codes must be consistent. In real life, there are always errors in this manually entered data: missing range, overlapping range, unit confusion (ppm and % mixing), different spellings of the lithology code (GRN, Grn, granite), codes such as 999 or -9 instead of a null value. AI is very powerful at scanning and flagging such inconsistencies; But you must make the correction with the approval of the geologist.
Step by step: drilling data cleaning with AI
- Introduce the data. Give the AI the column names, units, and a few sample rows (anonymizing the actual coordinates). Describe clearly what should happen.
- Make a list of rules. Ask the AI for a list of "what consistency rules should I check for this table", add your own rules on top.
- Have an inconsistency scan performed. Just ask the AI for flagging, not autocorrect. Ask for justification for each sign.
- Return to the raw source. Check each marked line against the original borehole log or laboratory certificate.
- Make the decision to fix it. The geologist makes the correction; AI just suggests.
Tip: Don't tell the AI "give me the corrected table"; Say "flag suspicious lines with reason". In an autocorrecting table, it is almost impossible to notice rows that are silently corrupted.
Core photo and visual interpretation
In modern projects, core boxes are photographed. Multi-model AI tools (models that can also process images) can produce a preliminary interpretation from these photographs: color change, possible fracture density, presence of veins, possible alteration zones. This is valuable as a preliminary screening that tells the geologist "look carefully here." However, lithology and alteration decisions are never made from photographs; Manual examination of the core, hardness test, acid reaction and laboratory analysis are required. It is a serious mistake for the AI to infer "there is gold" from the color; Gold is often invisible to the naked eye and can only be identified by analysis.
three mini cases
Case 1 — Unit interference capture. In a project, there are 12,000 rows in the assay table. The geologist introduces the table to AI and says, "List the rows with values above 100 g/t in the Au column; these may be either true high grade or ppm/ppb mixing." AI marks 37 lines. The geologist looks at laboratory certificates; He finds that 34 were actually entered as ppb (one per billion) but were mistakenly thought to be g/t, while 3 were the real bonanza (very high grade). If these 34 lines were not corrected, the average grade would be artificially inflated and the reserve would be exaggerated.
Case 2 — Overlapping range. In a lithology table, some of the from-to intervals overlap: 45.0-52.0 m of granite and 50.0-58.0 m of shale are defined simultaneously. AI lists these overlaps and shows a total of 19 lines of overlap. The geologist turns to the raw log; finds and corrects a line shift in data entry. If overlapping intervals were not corrected, it would remain unclear which rock type would be assigned to the block model and the tonnage would be calculated incorrectly.
Case 3 — Incorrect visual inference (warning). An intern gives the core photo to the AI and asks, "Is there any ore in this zone?" AI sees redness and veins and says "possible mineralization is high." The intern processes this into the log as "mineralized". Whereas the redness is iron oxide staining and does not contain economic ore; Laboratory analysis finds the grade at trace levels. Lesson: grade/ore cannot be determined from the image; The AI output is just an “examine” sign, verified by analysis.
Copiable prompt templates
DRILL DATA CONSISTENCY SCANNING "Role: You are an assistant resource/database geologist. Below is a [lithology/assay] table spaced from-to. Columns: [list, with units]. Mark the following inconsistencies with ROW number and reason: from>=to, overlapping range, gap, suspected unit mix-up, different spelling of repeated code, negative/outlier, code instead of blank (999/-9). DO NOT CORRECT, just mark Table: [paste].”
LITHOLOGY CODE STANDARDIZATION "The following lithology codes were entered by hand and are inconsistent. Group different spellings of the same rock type (e.g. GRN/Grn/granite). Write a single recommended standard code for each group. In a table: [original codes] -> [recommended standard]. Mark the groups you are unsure of separately; let me decide. Codes: [paste]."
CORE PHOTO PRE-SCANNING "List the visual features I should notice in this core photo: color change zones, possible fracture/fracture density, visible vein, possible alteration mark. Write these as a HYPOTHESIS; DO NOT decide on rock type or grade. For each feature, indicate what the geologist should check in the field."
GEOLOGICAL LOG SUMMARY "The following text is the geological log of a drilling. Turn it into a structured summary with the following headings: dominant lithologies and depth ranges, alteration zones, structural notes (fault/fracture), mineralization observations. Do not add anything that is NOT in the text; mark ambiguous phrases as 'unclear'. Log: [paste]."
Weak prompt / Strong prompt
WEAK PROMPT: "Fix this drilling data."
STRONG PROMPT: "Role: You are an assistant database geologist. In the assay table below, flag ONLY suspicious rows (from>=to, ppm/% suspected adulteration, negative value, 999 instead of blank). Give row number + reason for each flag. DO NOT autocorrect; I will check the correction from the raw certificate. Columns: HoleID, From(m), To(m), Au(g/t). Table: [paste]."
The weak prompt tells the AI "you decide and change"; The powerful prompt limits the AI to marking, leaving correction authority to the geologist.
Comparison table: How to use AI in which business
Quest
The role of AI
whose decision
verification
Inconsistency marking
Powerful browser
geologist fixes
Raw log/certificate
Lithology code merging
Suggestion
Geologist confirms
Code dictionary
Core photo comment
Pre-screening/hypothesis
Geologist/laboratory
Manual inspection + analysis
Grade/ore decision
Not used
laboratory
Accredited analysis
Summarize log text
configurator
geologist
Comparison with original text
Common mistakes
- Having the AI “autocorrect” the table. You may not notice silent distortions; just mark it.
- Grade/ore extraction from core photo. Gold is often invisible to the naked eye; The decision is your analysis.
- Pasting real coordinates without anonymizing. License and coordinates are sensitive data.
- Not comparing the AI summary with the original log. AI can add an observation that is not in the text (hallucination).
- Ignoring unit mismatch. ppm/%/g/t/ppb mixing directly disrupts the reserve.
Attention: A unit error in drilling data does not remain in a single line; Compositing grows and spreads throughout the forecast and reserve. Diligence in the cleaning phase determines the reliability of all subsequent accounts.
In summary
Drilling/geology data consists of four main tables (collar, survey, lithology, assay) and consistency rules are vital. AI is a powerful aid in this data inconsistency scanning, code standardization, log summarization and core photo pre-scanning. However, the correction decision belongs to the geologist, and the grade/ore decision belongs to the laboratory. Have AI mark not edit, verify each mark from the raw source, and never infer grade from the image.
Application task
Anonymize an assay table you have (or a sample) and give it to AI with the "Drilling data consistency scan" template. Import the marked rows into a table; Consider a "to be checked from raw source" note for each. Then create the code dictionary with the “Lithology code standardization” template and separately note the groups that the AI is unsure of. Finally, pre-scan a core photo and write down what you will check in the field for each hypothesis that emerges.
checklist
- [ ] I anonymized the actual coordinate/license information before giving the data to the AI.
- [ ] I asked AI for a justified marking, not a correction.
- [ ] I checked each line marked against the raw log/lab certificate.
- [ ] I also verified volume consistency (g/t, %, ppm, ppb).
- [ ] I considered the core photo output as a hypothesis, I did not make a grade decision.
- [ ] I compared the AI summary to the original log text to verify that there was no information added.