Universal 3D models are trained on a controlled corpus exceeding one million paired samples, augmented with edge cases derived from 300+ real-world projects. This broad geological and physical coverage enables downstream models to be interpretable in terms of training data exposure, rather than treated as black-box statistical outcomes.

Step 4: Universal Model Training & Provenance

Synthetic acquisition responses are generated using acoustic and elastic finite-difference modeling, including anisotropic wave propagation and angle-dependent reflectivity (e.g., Born approximation). The forward-modeled data are then imaged using multiple migration families—RTM, Kirchhoff, and one-way wave equation (OWWE). This workflow reproduces realistic acquisition and imaging artifacts such as illumination gaps, migration smiles, and subsalt shadowing.

Step 2: Wavefield Simulation & Migration Realism

Because geological objects are generated explicitly, target labels (fault masks, salt boundaries, channel bodies) represent exact ground truth and are free from interpreter bias. We perform multi-parameter dataset-level quality control and balancing across geometry, frequency content, noise levels, depositional settings, and imaging variants to ensure a transparent and well-conditioned training corpus.

Step 3: Exact Labels & Dataset-Level QC

We generate large ensembles of 3D structural scenarios (fault systems, salt bodies, channel–levee complexes) with explicit parameter control. Each realization is converted into a physically consistent elastic (or viscoelastic) earth model populated with VpVp​, VsVs​, density, attenuation, and anisotropy parameters. This links geological geometry with wave physics, ensuring realistic amplitude behavior, phase response, and interference patterns.

01 / AI-Powered Seismic Interpretation Engine

Step 1: Structural Scenario & Elastic Earth Model Generation

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

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

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

The platform operates directly on raw amplitude data. No pre-processing, no spectral decomposition, no seismic attribute calculation needed before inference. For paleo-channel and reef detection models, this eliminates the traditional dependency on dozens of secondary seismic attributes — removing a major source of interpreter time and methodology bias.

02 / Neural Network

Input: Raw Post-Stack Seismic Amplitude, No Pre-Processing Required

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

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

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

The geophysicist manually interprets just one or two 2D sections (inline/crossline) using full domain expertise. This minimal expert input becomes the single source of local geological ground truth. The challenge solved: a model trained on Gulf of Mexico deep-water data fails on North Bahrain sub-salt complexes or Latam sub-salt settings — Transfer Learning corrects this domain shift in about 20 minutes.

03 / Transfer Learning

Input: 1–2 reference cross-sections, manually interpreted