Gains:
- Ability to control artificial intelligence-based noise suppression with difference plot and prevent weak but real reflection from being erased
- Ability to recognize the risk of deconvolution producing false resolution and the trap of trace interpolation creating ghost structures and follow them with a mask
- Ability to verify physical validity by connecting processing results with a well tie
What you see when you open a raw seismic record is rarely a "picture of the subsurface." Instead, you see a complex mix: the actual signal reflected from underground, surface waves, random noise, multiple reflections, and inherent distortions of the recording. Seismic data processing is the process of step by step cleaning, aligning and transforming this raw record into an interpretable image of the underground. In this unit, we will explain how artificial intelligence (AI) is a powerful accelerator of this process, especially in steps such as denoising, deconvolution and missing trace filling; but we will see why parameter and physical validity decisions still remain with the expert.
Anatomy of seismic data
Let's start with a short glossary. A trace is a time-dependent amplitude series recorded by a single receiver. When many traces come together, a section is formed. The sample interval is the time between two consecutive amplitude values (e.g. 2 ms). The actual reflections within the signal arise from the acoustic impedance difference of the layer boundaries in the subsurface; The purpose of processing is to separate these reflections from noise and place them in the correct time-space location.
Noise comes in two main types: random noise (incoherent, independent of tracks) and coherent noise (coherent noise — regular, strong, masking the signal, such as surface waves/ground roll). AI-based methods are often more flexible than traditional filters in separating these two types because they use a complex discrimination learned from samples rather than a fixed frequency-speed window.
Noise suppression with AI
Classical noise suppression is done with filters based on frequency, speed (f-k filter) or statistics. AI-based noise suppressors (mostly convolutional neural networks, CNN for short; networks that learn local patterns in image-like data) learn from "noisy input → clean output" sample pairs. Its advantage: it can make a context-sensitive distinction instead of a fixed rule when separating signal from noise. Risk: the tendency to delete the signal or leave the noise as a signal when encountering a type of noise that is not in the training data.
The most critical danger is this: an aggressive AI noise suppressor can mistake a faint but real reflection for noise and eliminate it. So the golden rule: be sure to view the difference (removed component) before and after noise suppression. If you see a consistent reflection pattern in the removed part, the filter is eating signal.
Tip: Plot the "noise" layer produced by the YZ noise suppressor as a separate slice (difference plot). The residue of a good repressor appears randomly; If there are slanted, continuous reflections in it, the filter is too aggressive and you need to soften the parameter.
Deconvolution: getting back the source signature
Deconvolution is the process that attempts to reveal the true reflectance series of the subsurface by subtracting the wavelet signature of the source (source wavelet — the waveform produced by the blast/vibrator) from the recorded trace. The goal is to increase resolution and suppress multiple reflections. AI can improve wavelet estimation and robustness to noise in deconvolution; However, if decon is applied excessively, it will amplify high-frequency noise and produce spurious resolution. AI is a recommendation engine here too; The final parameter selection (decon operator length, pre-whitening) is expertly calibrated according to the physical character of the traces.
Incomplete trace interpolation and regular sampling
Some receivers fail in the field, some lines are collected incompletely; As a result, gaps occur in the data. AI-based interpolation (filling in missing traces by learning from neighbors) is very useful in obtaining a regular grid before migration. However, the filled trace is an estimate, not a measurement; These traces should not be treated with equal weight as the real data in the interpretation, but should be marked separately in the output.
Caution: If traces produced by interpolation proceed as "real data" in subsequent steps, they may lead to the creation of a non-existent structure. Trace the filled cells with a mask; If a critical structure appears only in the interpolated region, do not trust it.
three mini cases
Case 1 — Network learning ground roll. In land seismic, strong surface waves (ground roll) were masking the signal. The team used a CNN noise suppressor that learned from 800 hand-cleaned recordings; processing time decreased from 12 minutes to 40 seconds per record. The team manually inspected 5% recordings from each batch and confirmed there was no signal loss with a difference plot.
Case 2 — Deleted weak reflection. In another project, an aggressively tuned YZ suppressor mistook a faint but real reservoir reflection at 2.4 seconds for noise and erased it. The difference only became apparent when the well-seismic tie was established: the clear layer in the well was not present in the section. When the parameter was smoothed and reprocessed, the reflection returned.
Case 3 — The ghost of interpolation. In a 3D volume where missing lines were filled in by AI, the commentator noticed a subtle channel structure. Looking at the fill mask, he saw that the channel fell exactly into the interpolated space. When additional field data was collected, it became clear that the channel was not there, but a pattern produced by the forecast.
Four copyable templates
1) Processing flow sketch:
Your role: seismic processing specialist. Data: [land/sea, 2D/3D, sampling 2 ms, target 1-4 km]. Outline me a step-by-step processing flow (pre-processing → noise suppression → decon → velocity analysis → migration). In EACH step, write down which decision requires parameters and what I should look for in choosing that parameter.
2) Noise suppression control plan:
I applied an AI noise suppressor. Give me an audit checklist: what should I look for in the difference plot, how do I recognize signal loss, what metric should I compare before/after, what recording rate should I manually check? Write short and actionable.
3) Deconvolution parameter consultation:
Your role: processing consultant. I apply deconvolution and high frequency noise increased. Explain possible causes, operator length, and the effect of the pre-whitening setting. How do I know if excessive deco is producing false resolution? Give reasons for each suggestion.
4) Interpolation mask tracking:
I filled in the missing traces with AI. Tell me how to keep the filled cells marked in the next steps, how to separate the comment from the actual data, and how to eliminate the "structure visible only in the interpolated region".
Weak prompt / Strong prompt
Weak prompt:
Clear the noise in seismic efficiency, tell the best setting.
A “best setting” does not exist independent of data; AI gives a general, uncontrollable recommendation.
Powerful prompt:
Your role: seismic processing specialist. Yield: black 2D, strong groundroll (5-12 Hz), target reflections 20-45 Hz, sampling 2 ms. Task: propose an approach for noise suppression, explain in frequency-rate terms which component to preserve and which to suppress. How do I monitor the risk of signal loss (differenceplot metrics) and with what metric do I compare before/after? Don't just say "enter this setting"; Give decision criteria.
Given the data character, target band, and control criteria, the output becomes truly feasible.
Processing steps and AI role
Processing step
AI contribution
Main risk
human decision
Random noise suppression
high
Delete weak signal
Parameter confirmation
Consistent noise (ground roll)
high
excessive suppression
Difference plot control
deconvolution
medium
fake resolution
Operator/whitening
Trace interpolation
medium
ghost building
Mask tracking
Speed analysis
medium
Migration at wrong speed
Confirmed by well
Common mistakes
- Suppressing without looking at the difference plot. You won't notice signal loss until you see the removed component.
- Overdoing the decon. Mistaking high-frequency noise for resolution produces spurious layers.
- Count the interpolated trace as real data. Leaving filled cells unmarked creates a ghost structure.
- Looking at one metric. It would be misleading to look only at the signal-to-noise ratio and ignore reflection conservation.
- Bypassing the well tie. Moving on to interpreting the processing result without connecting it to the well data.
In summary
In seismic processing, AI provides greater speed and often more flexible separation in noise suppression, deconvolution, and trace interpolation. But there is a physical pitfall at every step: erasing weak signal, producing false resolution, giving rise to ghost artifacts. That's why controls such as difference plot control, well tie and interpolation mask are indispensable. AI suggests parameters; It is the expert who validates the parameter according to the physics of the data and makes the result interpretable.
Application task
Define a noise suppression step on a seismic section (real or sample). Extract the flow with the "Processing flow outline" template, then produce a checklist with the "Noise suppression control plan" template. Compare the before/after slices and the difference plot and evaluate in your own words whether there is consistent reflection in the removed component.
checklist
- [ ] I compared before/after noise suppression and difference plot.
- [ ] I have verified that there is no consistent signal in the removed component.
- [ ] I checked the decon result for false resolution.
- [ ] I kept the interpolated traces marked with a mask.
- [ ] I tied the result with the well data (well tie).