Every topic, alphabetical

Definition coming soon.

Definition coming soon.

Materials whose architecture mimics natural design — nacre's brick-and-mortar toughness, spider silk's strength-to-weight ratio, lignin's cross-linked aromatic scaffold. The category covers both literal copies and the broader principle: structure earned over evolutionary time tends to beat structure assumed by an engineer. AI accelerates the translation by searching the formulation space orders-of-magnitude faster than wet-lab iteration.

Organic matter from plants, agricultural residues, forestry waste, and pulp-mill side streams — abundant, renewable, geographically distributed. The energy transition's underused feedstock category: chemistry-rich, AI-tractable, and politically easier to deploy than new offshore wind. EarthScan's MIMEA work targets the prediction layer that turns heterogeneous biomass into a priceable input for downstream industries.

Definition coming soon.

Neural networks that learn local spatial features through convolution operations — the workhorse architecture for image and grid-structured data since AlexNet. Underpins every dense-prediction subsurface task: seismic facies classification, fault detection from 3D volumes, raster-log curve segmentation. Modern variants combine CNN feature-extraction with transformer-based long-range attention.

Definition coming soon.

Definition coming soon.

The infrastructure that moves raw data from source systems (sensors, well-archives, partner APIs) through cleaning, transformation, validation, and feature engineering into model-ready feature stores. The unsexy 60% of any production-AI project — and where most upstream-AI pilots actually die. EarthScan's MLOps work emphasises pipeline robustness as a precondition to model accuracy.

The school of thought — championed by Andrew Ng — that says: hold the model fixed, iterate on the *data*. Most production-grade gains in subsurface AI come from cleaner labels, smarter sampling, and better-curated edge cases — not bigger transformers. EarthScan's pipelines are built data-first: every drop in error gets traced back to a label-quality, coverage, or distribution-shift fix before any architecture change is considered.

A subset of machine learning where neural networks with many layers learn hierarchical representations directly from data. The dominant approach in modern subsurface AI: every published EarthScan model — VeerNet, EarthAdaptNet, the Hamiltonian Monte Carlo reservoir-characterisation work — uses deep learning at its core.

Definition coming soon.

A family of techniques that lets a model trained on one labelled dataset (the source domain) work well on a different unlabelled dataset (the target domain). Critical for subsurface AI because labelling seismic volumes is expensive — domain adaptation lets you annotate one basin and deploy across many. EarthScan's EAN-DDA uses CORAL (correlation alignment) at the encoder bottleneck for unsupervised cross-basin transfer.

Definition coming soon.

EarthScan's encoder-decoder neural network architecture for seismic-facies semantic segmentation. Built on Residual Blocks (contracting path) + Transposed Residual Blocks (expanding path), with U-Net-style skip connections at every depth. The EAN-DDA variant adds CORRELATION ALIGNMENT (CORAL) for unsupervised deep domain adaptation between basins. Published at IEEE TGRS 2022.

A neural network architecture pattern where the encoder progressively reduces input spatial resolution while increasing feature depth, and the decoder mirrors that path back to original resolution. The dominant shape for dense-prediction tasks (segmentation, depth, super-resolution). Both VeerNet (raster-log digitisation) and EarthAdaptNet (seismic-facies segmentation) use encoder-decoder backbones with U-Net-style skip connections.

The shift from fossil-fuel-dominated energy production toward renewable, low-carbon, distributed sources — solar, wind, geothermal, biomass, hydrogen. Subsurface AI plays in two ways: (1) accelerating fossil-fuel decline by getting more value out of existing assets, and (2) making renewable feedstocks (geothermal reservoirs, biomass streams) commercially predictable through prediction layers.

Definition coming soon.

A public 3D seismic dataset from the Dutch sector of the North Sea, widely used as the canonical benchmark for academic seismic-AI work. Released by dGB Earth Sciences with annotated horizons and facies. EarthScan's EAN-DDA paper trained on F3 (source domain) and adapted to Penobscot (target) to demonstrate cross-basin transfer.

Identifying tectonic faults in 3D seismic volumes — critical for both geohazard assessment (drilling stability, wellbore kicks) and hydrocarbon exploration (faults can be either seals or conduits). Traditionally a multi-day senior-geophysicist task; modern deep-learning approaches segment fault probabilities at the voxel level. The unsolved part is fault-continuity — turning per-voxel probabilities into clean connected fault surfaces.

Definition coming soon.

Identifying natural and induced fractures from borehole images and well-log data. Fractures are diagnostic for both production (planning hydraulic stimulation) and wellbore stability (detecting weak zones before they collapse). FMI (Formation Micro-Imaging) logs are the typical input; EarthScan applies CNN-based unsupervised segmentation to detect bedding planes and fracture sinusoids without labelled training data.

Definition coming soon.

Definition coming soon.

Definition coming soon.

Definition coming soon.

Definition coming soon.

Definition coming soon.

The second-most-abundant biopolymer on Earth — what holds wood together. Roughly 100 million tonnes per year are produced as a waste stream from paper and cellulosic-biofuel manufacturing, but less than 2% gets sold as a structural input. Heterogeneous, branched, polyaromatic — the heterogeneity that makes lignin commercially under-used is exactly where AI gets a real edge.

The objective the optimiser minimises — and the single most underrated lever in applied deep learning. Cross-entropy is fine for balanced classification; segmentation with rare classes wants Dice or Lovász-Softmax; ranked retrieval wants triplet or contrastive. EarthScan's segmentation work leans on Lovász-Softmax for the IoU-aligned gradient signal it gives on geophysics-typical class imbalance.

Definition coming soon.

Definition coming soon.

The umbrella discipline behind every algorithm that learns patterns from data rather than being explicitly programmed. Encompasses classical methods (random forests, SVMs, regression) plus the deep-learning subset that dominates modern AI. EarthScan uses the full toolkit — pragmatic about which algorithm fits a given subsurface task.

The discipline that connects atomic structure to bulk properties — the reason a polymer blend is stiff or a composite resists fatigue. AI shifts the historical bottleneck: instead of fabricating-then-testing every candidate, models trained on density-functional-theory + experimental corpora propose viable formulations before a beaker is touched. EarthScan's MIMEA programme sits exactly here, scoped to bio-derived feedstocks.

The plumbing that turns a notebook experiment into a production system — versioned data, reproducible training, monitored inference, automated rollbacks. Subsurface ML lives or dies on this layer: a model that drifts on a new basin without anyone noticing is worse than no model at all. EarthScan treats MLOps as a first-class deliverable, not an afterthought.

Definition coming soon.

Definition coming soon.

Public-domain 3-D seismic dataset off the Scotian Shelf, Canada — released by the Nova Scotia Department of Energy and widely used as a community benchmark for facies classification, fault detection, and self-supervised pre-training. Smaller and structurally simpler than the F3 Block, but cleaner — useful as a sanity-check companion to F3.

Definition coming soon.

The process of converting scanned paper well-log images into clean, queryable digital curves (LAS format). Decades of legacy logs in mature basins exist only as raster archives — manual digitisation runs hours per log with high inter-interpreter variance. EarthScan's VeerNet AI model is the production pipeline: a U-Net + transformer hybrid that turns raster images into digitised curves at ~50× the manual throughput.

Definition coming soon.

The discipline of describing a reservoir's geometry, rock properties, fluid distribution, and behaviour over time — the input every drilling, completion, and production decision rests on. EarthScan's tools converge here: VeerNet recovers the well-log curves, EarthAdaptNet maps facies in seismic, downstream models stitch them into property cubes that drive plays.

Definition coming soon.

Definition coming soon.

Classifying every voxel in a seismic volume into one of N geological facies (sand / shale / channel / salt / etc.). The cornerstone of automated subsurface interpretation — what was a senior-geophysicist-week becomes minutes once the model is deployed. EarthScan's EAN-DDA architecture performs this on the F3 (Netherlands) and Penobscot (Canada) public datasets with demonstrated cross-basin transfer via deep domain adaptation.

The set of geophysical techniques that turn surface-recorded reflection data into a 3-D picture of the subsurface — migration, velocity modelling, depth conversion, attribute extraction. The image is the input that every downstream EarthScan model consumes; image quality bounds model quality. Recent AI work targets the whole stack from denoising and demultiple to direct property inversion.

The task of classifying every pixel (or voxel) in an image into one of N categories. The technical foundation under most subsurface-AI workflows: seismic facies analysis is segmentation across geological classes; fault detection is segmentation between fault and not-fault voxels; raster-log curve identification is segmentation between curve traces and background. EarthAdaptNet was designed specifically for this on seismic volumes.

Definition coming soon.

Definition coming soon.

Definition coming soon.

Definition coming soon.

Definition coming soon.

Definition coming soon.

The architecture that replaced recurrence with attention and ate the rest of deep learning in the process — language, vision, audio, and increasingly the geosciences. Subsurface-flavoured transformers (SeisBERT, SubsurfaceGPT, SAM-style segmenters) bring long-range context to seismic, well-log, and core-image tasks where convolutions plateau. EarthScan uses them where they earn their compute, not by default.

The encoder–decoder architecture with skip connections that became the default backbone for pixel-wise segmentation across medicine, satellite imagery, and — for EarthScan — seismic facies. The skip connections are the trick: they let the decoder recover spatial detail lost in the bottleneck. EarthScan's segmentation pipelines run modified U-Nets with residual blocks and Lovász-loss for the geophysics edge cases where vanilla cross-entropy misbehaves.

Definition coming soon.

Definition coming soon.

The deep-learning model EarthScan built specifically for raster well-log digitisation. Encoder-decoder architecture with residual blocks in the contracting path, transformer self-attention at the bottleneck for long-range curve dependencies, and bilinear-upsampling decoder blocks that reconstruct per-pixel curve masks. Published in MDPI Journal of Imaging 9(7), 136 (2022). Powers the ES Raster Digitizer product line.

Definition coming soon.

Definition coming soon.

Continuous geophysical measurements taken at depth in a borehole — gamma ray, resistivity, density, neutron porosity, sonic, caliper, and many more. The single richest data source in upstream subsurface analysis. Modern well-logs are recorded digitally as LAS files; the legacy archive is millions of paper / raster scans that EarthScan's VeerNet pipeline turns back into digital curves.

Definition coming soon.