Earth_model → Vp, Vs, ρ, impedance, Q, anisotropy
Linkage → Geometry × facies × petrophysical priors
Result → Physically interpretable amplitude behaviour

Each structural realization is converted into a physically usable earth model. We populate layers and objects with Vp, Vs, density, impedance contrasts, attenuation and, where required, anisotropy parameters. This is the critical bridge between geometry and wave physics: without a plausible elastic model, synthetic seismic will look seismic-like but will not teach the network the right amplitude and interference behaviour.

Step 02 — Rock Properties
and Elastic Earth-Model Population

Methods → Acoustic_FD, elastic_FD, TTI/anisotropic_modelling, reflectivity/Born-style approximations
Physics → AVO/AVA, multiples, diffractions, tuning, attenuation, illumination variation
Noise → Coherent, random, migration-related, acquisition footprint

We then generate synthetic acquisition responses with multiple modelling engines depending on the target problem. The toolkit includes acoustic and elastic finite-difference modelling, anisotropic propagation where needed, reflectivity-based approximations for rapid coverage expansion, and controlled noise injection. The aim is not only realism, but controlled diversity in wavelet, bandwidth, phase, illumination and noise regime.

Step 03 — Wavefield Simulation

Imaging → RTM, Kirchhoff, one_way_wave_equation, stack-domain variants
Teaches → Illumination gaps, migration smiles, shadow zones, positioning uncertainty
Benefit → Model learns field-like seismic, not idealized wavefields

A major part of the realism comes from the imaging stage. We generate training data not only at the forward-model level, but also after seismic imaging with different migration families. Depending on the use case, the pipeline can include RTM, Kirchhoff migration and one-way wave-equation imaging, so the model sees realistic illumination gaps, migration swings, sub-salt shadowing and other artefacts that interpreters actually face on field data.

Step 04 — Imaging and migration realism

Pairing → Seismic_tensor: S, label_tensor: L, provenance: P
QC → Class_balance, geometry_balance, frequency_balance, imaging_balance
Trace → Every sample linked to generation assumptions

Because every object in the model is generated explicitly, the labels are mathematically exact rather than interpreter-dependent. Fault masks, salt boundaries, channel bodies and structural surfaces are paired directly with the simulated seismic response. We then run dataset-level quality control: geometry balance, frequency balance, noise balance, basin-style balance and imaging-variant balance. This is the core of training transparency.

Step 05 — Exact labels and dataset QC

Base_model.train (samples=1_000_000+, exposure=GLOBAL)
Coverage → Tectonic_styles × imaging_variants × noise_regimes
Moat → Data provenance + exact labels + controlled realism

Only after the corpus has geological and physical coverage do we train the universal 3D models. The result is not merely scale, but coverage with provenance: more than one million paired samples, 300+ real projects worth of accumulated edge cases, and a dataset whose composition is understood. That is why the downstream models are explainable in terms of training exposure rather than treated as opaque statistical luck.

Step 06 — Universal model training on the controlled corpus

encoder → [gradients, textures, discontinuities, phase_patterns]
dilated → receptive_field × N² | compute_cost × 1
result → macro_context + local_resolution simultaneously

The encoder progressively extracts features at multiple spatial scales. For the fault detection model, dilated convolution blocks are integrated between encoder modules — exponentially expanding the receptive field without increasing computational load. This allows the network to simultaneously analyse macro-geological context (regional fault systems) and local fault micro-throws with no resolution trade-off.

Encoder — hierarchical feature extraction with dilated convolutions

kinematic_preservation → TRUE | dynamic_preservation → TRUE

skip_connections → HF_components → decoder (bypass bottleneck) SNR_improvement → 1.4× – 5.2× (validated on real industrial projects)

This is the core reason for choosing Residual U-Net specifically for geophysics: skip-connections carry high-frequency spatial information — sharp reflector boundaries and fault planes — directly from encoder to decoder, bypassing the downsampling bottleneck. Classical structure-oriented smoothing filters destroy this information when suppressing noise aggressively. Our architecture preserves it. Result: SNR improvement of 1.4× to 5.2× with full kinematic and dynamic signal preservation — critical for subsequent AVO inversion.

Decoder + skip-connections — the key to vertical resolution preservation

output → probability_volume[fault | channel | salt | reef] method → no window_sliding, no base_summation → zero ringing artefacts
post → geobody_extraction + surface_generation + volumetric_filter

All detection models output continuous probabilistic volumes — the geophysicist controls the confidence threshold. The inference pipeline detects paleobodies in the 3D space at their true time/depth position — without window-sliding techniques that introduce ringing artefacts from surrounding "singing" objects. Post-inference: automatic geobody extraction, structural surface generation, and intelligent volumetric filtering.

Output — continuous probability volumes, not binary masks

frozen_layers → early_encoder (universal_seismic_features)
frozen_weights → gradients, textures, discontinuities,
phase rationale → 1M+ samples of knowledge must not be overwritten

Early encoder layers learn universal, domain-agnostic seismic features: amplitude gradients, phase patterns, reflection discontinuities, texture variations. These weights represent the accumulated learning from 1M+ training samples and are frozen — they must not be overwritten by the small local dataset. This is the practical guarantee of stability: the network does not forget general seismic physics while adapting to local tectonic style.

Freeze — preserving 12 years of universal seismic knowledge

fine_tune (layers=decoder_final, data=local_sections)
learns → [local_tectonic_style, fault_geometry, salt_morphology]
ignores → migration_smiles (blur_zone_preprocessing)
time → ~20 min on modern GPU

Only the final decoder layers are fine-tuned on a modern GPU to match the user's specific interpretation geometry. The network learns the local tectonic style: complex flower structures, shear zones, salt diapir morphology and thin-bed interference patterns. It also learns to de-emphasize migration artefacts flagged by the interpreter as noise, while preserving the broader physical priors captured during base training.

Fine-tune — adapting to local tectonic geometry in ~20 minutes

model.infer (volume=FULL_3D, any_size=TRUE, any_vintage=TRUE)
accuracy → expert_validated
dataset_prep_time → ~0
overfitting_risk → LOW (frozen encoder prevents memorisation)

The re-calibrated model runs inference on the entire 3D volume. Expert-validated, basin-specific results regardless of survey area, vintage, noise level or tectonic complexity. Near-zero time spent building local labelled datasets. The approach mathematically guarantees high-quality results: the frozen universal weights prevent overfitting to the two reference sections, while the fine-tuned decoder ensures geological coherence across the full volume.

Inference — expert-validated 3D volume, any size, any complexity