Gains:
- Ability to understand potential field processing corrections (Bouguer, land, log, IGRF) and the parameter sensitivity of the regional-residue separation and test it with different separations
- Ability to interpret source location correctly, taking into account depth-amplitude uncertainty and magnetic latitude effect (reduction to pole)
- Ability to cross-validate potential field interpretation with alternative scenarios and seismic/well without locking the potential field interpretation to a single model
The seismic method is expensive and cannot be applied everywhere. This is where potential field methods—gravity and magnetic—come into play: relatively inexpensive methods that quickly scan large areas and reveal the rough structure of basins and ore bodies. The gravity method measures very small gravitational changes produced by density differences underground; The magnetic method measures field changes arising from differences in magnetization (magnetic susceptibility) of rocks. In this unit, we will explain how artificial intelligence (AI) accelerates the processing of these two methods, anomaly discrimination and depth estimation; We will see why the strong uncertainty inherent in potential fields requires using AI with extra caution.
Nature of potential field data
Let's start with a dictionary. Anomaly, deviation from the expected regional value; In other words, it is the signature of the underground structure we are interested in. In gravity, the raw measurement undergoes many corrections: free-air correction (effect of elevation), Bouguer correction (effect of rock mass between the measurement and reference) and terrain correction (effect of topography). The resulting Bouguer anomaly reflects subsurface density differences. In magnetics, the daily variation and the main field (IGRF) are subtracted to obtain the magnetic anomaly.
The main challenge of potential fields is the depth-amplitude uncertainty: a small and shallow source and a large and deep source can produce almost the same surface anomaly. This is the potential field state of polysemy in seismic, and AI doesn't eliminate it — it just accelerates it. That's why in commentary I always ask "how many different models explain this anomaly?" question is asked.
AI in processing and field separation
The heart of potential field processing is regional-residual separation: the separation of the slowly varying regional component from broad, deep sources and the residual component from shallow, local sources. Classical methods are polynomial fitting or wavelength filters; AI-based approaches can make this distinction more flexible by learning from examples. Additionally, derivative-based features—horizontal/vertical derivatives, analytical signal, tilt angle—highlight source edges; AI can combine these attributes and automatically mark edge and resource candidates.
The risk is that the filter parameter (cutoff wavelength) completely changes the interpretation. The same data produces completely different residual maps with different distinctions. Even if the AI suggests a separation, the expert checks whether that separation is suitable for the target depth.
Tip: Do not fix the regional-residue distinction with a single parameter. Generate three or four residual maps with different cutoff wavelengths and see if the target anomaly looks consistent on all of them. An anomaly that only appears in a single distinction is often a parameter artifact.
Depth estimation and automation
Methods such as Euler deconvolution and Werner deconvolution are used to estimate source depth in potential areas; these estimate the source location and depth from the gradients of the anomaly. AI can speed up these predictions and produce a cleaner “depth cloud” by clustering thousands of solutions. However, these methods are very sensitive to inputs such as the structural index — the assumption that defines the geometry of the source; False assumption systematically yields false depth. The depth map produced by the AI is always inspected by geological plausibility and seismic/well if possible.
Caution: Due to polarity and latitude effects in magnetic data, the anomaly may not be directly above the source. In the northern hemisphere, the anomaly may shift south of the source. Do not accept the "resource center" marked by AI where it is without reduction to pole.
Interpretation integrated with seismic
Potential fields alone are of poor resolution; Their real power emerges in joint interpretation. Gravity gives the depth of a basin, magnetic gives the topography of the base, and seismic gives the fine structure. AI helps combine these different data layers into a common model (data fusion); but the compositing is done under expert supervision, taking into account the resolution and uncertainty of each method.
three mini cases
Case 1 — Narrowing the basin edge. In a geothermal project, it was necessary to find the basin edge (fault step) with gravity data. AI-assisted edge detection (tilt angle + analytical signal) marked 12 candidate edge lines. The team inspected these with magnetic and two existing wells; 5 were consistent with the basement fault step, 7 were noise-induced. The verified edge shifted the well location by 800 m.
Case 2 — Latitude effect fallacy. In a mineral exploration, the center of the magnetic anomaly was chosen as a direct drilling target. The first well turned up empty. Later, when reduced to the pole, it was understood that the real source was 300 m north of the anomaly center: the latitude effect was ignored. The second shaft found the ore body.
Case 3 — Parameter dependent ghost. One team saw a distinct "embedded structure" in the residual map produced by a single cutoff wavelength. When the map was reproduced with different cutoff values, the structure was lost: this was a parameter artifact of the regional-residue distinction, not a real source.
Four copyable templates
1) Machining correction control:
Your role: gravity/magnetic machining specialist. Yield: [land gravity,terrain rugged]. List me in order the corrections I need to apply (free air, Bouguer, terrain; daily +IGRF on pickup). For each correction, write the typical error and its effect in the comment if omitted.
2) Testing regional-residue discrimination:
I'm doing regional-residual discrimination, my target depth is ~[X] km. How do I choose the cutoff wavelength? Give me a procedure to compare residual maps produced with different cutoff values: how do I distinguish which anomaly is real and which is a parameter artifact?
3) Depth estimation control:
I estimated depth with Euler deconvolution. Explain how my choice of structural index affects the outcome. How can I compare the depth distributions obtained with different indices and confirm them with geological plausibility and seismicity?
4) Magnetic welding position correction:
Before making the center of my magnetic anomaly a drilling target, explain the latitude effect and the need to reduce it to the pole. How do I account for the offset between the source's actual location and the anomaly center? Tell me step by step.
Weak prompt / Strong prompt
Weak prompt:
Tell me where the ore is on this gravity map.
The potential area alone does not indicate the ore; AI hides uncertainty and produces single-point false certainty.
Powerful prompt:
Your role: potential field commentator. Context: Bouguer anomaly map, target sulphide ore, two wells present. Task:(1) mark candidate resource boundaries with edge detection attributes; (2) give at least two density-depth scenarios for each candidate, shallow-small and deep-large; (3) Tell me what additional data (magnetic, seismic, well) you recommend to separate these scenarios. Don't give a single precise resource location.
Uncertainty scenarios and cross-data requests make potential field interpretation honest and defensible.
Method comparison
feature
gravity
magnetic
AI contribution
The difference it measures
density
magnetization
Attribute merging
typical target
basin, salt, ore
Bedrock, ore
Edge/weld marking
Main uncertainty
Depth-amplitude
Latitude/pole effect
narrow down uncertainty
critical correction
Bouguer, land
Daily, IGRF, RTP
automation, control
Common mistakes
- The only difference is to trust the map now. Do not declare a structure without testing the anomaly with different cutoff values.
- Bypassing the latitude effect. The magnetic anomaly center is not directly above the source.
- Choosing the structural index arbitrarily. Euler depth is very sensitive to this assumption.
- Considering the potential area alone as definitive. Low resolution; Combine with seismic/well.
- Settling for a single model that hides uncertainty. Be sure to create shallow-small vs deep-large scenarios.
In summary
Gravity and magnetic methods cheaply scan large areas and provide a rough outline of the basin and ore structure; AI accelerates rendering, edge/source marking, and depth estimation. But the depth-amplitude uncertainty, latitude effect, and parameter sensitivity inherent in potential fields make the AI output extremely fragile. Correct usage: generating multiple scenarios, testing with different parameters, observing physics such as reduction to pole and combining with seismic/well. The final interpretation belongs to the expert who does not hide the uncertainty.
Application task
Select a gravity or magnetic anomaly (real or sample). Create a comparison routine for different cutoff wavelengths with the "Regional-residue discrimination test" template. If magnetic, evaluate the latitude effect with the "Magnetic source location correction" template. Write down at least two different subsurface scenarios that explain the anomaly and additional data to distinguish them.
checklist
- [ ] I applied all rendering corrections (Bouguer/terrain or log/IGRF).
- [ ] I tested the regional-residual distinction with different parameters.
- [ ] I took into consideration the latitude effect/reduction to pole in magnetics.
- [ ] I have produced at least two depth-amplitude scenarios for the anomaly.
- [ ] I cross-validated the result with seismic/well/other method.