White
papers

Long-form research from EarthScan. Peer-grade methodology, free with email. We publish a few of these a year — when there's something substantive to say.

Whitepaper
03 Jun 2026
18 pages · 19 min read
Why General LLMs Fail at the Wellbore: The Case for Domain-Native AI in Petroleum Geomechanics
General-purpose LLMs like Copilot scored 25% on a controlled petroleum geomechanics benchmark, fabricating numbers in two-thirds of data-grounded queries. Domain-native AI architectures — built SQL-first, module-scoped, and refusal-aware — scored 95.8%. This whitepaper presents the benchmark, the failure modes, and the engineering principles that close the gap.
Whitepaper
29 May 2026
28 pages · 18 min read
Hulde at the Wellbore: Domain-Native AI That Earns Engineering Trust
A domain-native language model purpose-built on petroleum geomechanics scores 95.8% across 14 live tasks versus 25% for general-purpose LLMs. When a mud-weight decision carries eight-figure NPT risk, the difference is not a product feature — it is an engineering safety margin.
Whitepaper
23 May 2026
18 pages · 15 min read
Own Your AI Stack: A Whitepaper for Energy Operators
AI in oil and gas will hit $17B by 2030, yet half of enterprise pilots never ship. The fix is engineering execution, not better models.
Whitepaper
20 May 2026
19 min
Running AI R&D Like a Board-Grade Program: RAID Logs, GDPR Rails, and Three-Audience Reporting
A multi-year AI research programme with a major operator in Oman survived long enough to ship because of the governance wrapped around the models, not only the models themselves. This whitepaper reconstructs that governance operating model from the engagement archive and reads it as the real risk-management instrument of the programme. The spine is a monthly cadence: daily stand-ups, weekly meetings, and monthly signed reports, all filing into a single RAID log that sorted every item into Risks, Actions, Issues and Decisions and covered both personal and wells data under one discipline. Sitting a full year ahead of that cadence was an Article-28-style data-processing agreement, signed 16 July 2020 in Assen under Dutch law, that fixed the EU-servers-only rails before any confidential log reached a model, a point an earlier chapter of this programme owns in full and we treat here as a pointer. The reporting itself was layered for three audiences at once: domain experts in geology and geophysics reading model-versus-ground-truth evidence, CIO and CDO leadership reading a before-and-after-MLOps maturity picture across eight dimensions, and the CEO, managing director and board reading a Three Horizons growth chart. And the programme was honest about its binding constraint: slow data collection was named the top risk, mitigated by a hard Phase-3 floor of 10 to 15 usable wells, a floor the delivered data cleared only after two of ten received wells were excluded for abnormal static value ranges, leaving eight usable. The claim throughout is that a research programme stays legible, auditable and fundable because of the cadence, the rails, the audience-layered reporting and the named risk, not because the models were good on any single week.
Whitepaper
05 May 2026
22 pages · 16 min
VeerNet: AI for Raster Well-Log Digitisation
Roughly 60% of the world's producing wells have legacy well-log data trapped behind raster scans that nobody has the time, money, or patience to digitise by hand. VeerNet — published in MDPI Journal of Imaging, 9(7), 136 — is the deep-learning architecture EarthScan built specifically to fix that. Hybrid CNN-residual encoder, transformer self-attention bottleneck, upsampling decoder. F1 = 35%, IoU = 30% on a 10K-image benchmark across five candidate loss functions. Now in production behind ES Raster Digitizer at multiple major operators with ~50× throughput vs interpreter-only baselines. This whitepaper covers the problem, the architecture, the loss-function ablation, the deployment, and what's next.
Whitepaper
03 May 2026
24 pages · 18 min
MIMEA: AI for Biomass and Bio-Inspired Materials
Lignin is the second-most-abundant biopolymer on Earth and the least commercially understood. This whitepaper makes the case for deep learning as the missing layer between heterogeneous biomass feedstock and high-value downstream chemistry. We cover Higher Heating Value prediction without a calorimeter, neural surrogates for atomic-scale potential energy surfaces, the multimodal data infrastructure required, and a sequenced 12-month adoption path. Drawn from the MIMEA feasibility study (Real AI B.V., 2021) and updated with the state of the art in 2026.
Whitepaper
15 Apr 2026
18 pages · 16 min
Lead KPIs, Lag KPIs, and a Bi-Weekly Verification Cadence: Governing an AI R&D Programme for Handover
An applied-AI research programme has three governance documents. The model write-up describes what was built; the invoice schedule describes how it was funded; the KPI contract describes what had to be measured for the money to move. The first two are published often. The third is the one that decides whether an open-ended research bet stays honest, and it is rarely read. We reconstruct the KPI contract behind a three-phase, roughly twenty-month subsurface computer-vision programme built for a major operator in Oman. Every phase carries a lead KPI and a lag KPI: Phase 1 pairs Time-to-Deploy within the phase window and a Standardization measure (a complete, ready dataset for the selected wells) as leads against reduced Time-to-Market as the lag; Phase 2 is gated on a contractual minimum of at least one fully functioning algorithm and at least one fully functioning model, versioned, with accuracy sound enough to carry into Phase 3; Phase 3 names self-service inference, tested, as its lead indicator, with model validation, operations and performance on distributed and HPC compute as the lags. Between Phase 2 and Phase 3 sits an intermediary handover report that is itself a contractual gate: it had to carry an infrastructure-cost advisory and an API-performance budget covering call duration and latency, so the running cost of the next phase was priced in writing before that phase was entered. A continuous bi-weekly progress cadence with validation, verification and preventive controls ran underneath all of it. The five payment tranches were wired to the gates, Part-1 due at signing, and the production and commercial phases (four and five) were held explicitly out of scope until a final recommendation. We argue that this KPI-to-gate-to-tranche wiring is the actual value-governance instrument of the programme: money follows proven capability rather than effort, and later commercial phases stay out of scope by design until the research proves out. Three interactive instruments walk the argument, and the effort-obligation clause work and the cash-flow ledger of this same programme are pointers here, not re-derivations.
Whitepaper
24 Mar 2026
19 min
Anatomy of a Multi-Phase, 20-Month Applied-AI Engagement: How the Money and Milestones Were Structured
A multi-phase applied-AI research contract is usually remembered for its deliverables. The document that decides whether the programme is fundable is the invoice schedule, and it is rarely read. This whitepaper reconstructs the cash-flow architecture of one such contract: a three-phase, roughly twenty-month engagement with a major operator in Oman, approved at a total of 533,000 US dollars. We trace the money along two axes. Along time, the total is paid as five milestone-gated tranches, 130,000, 104,000, 104,000, 104,000 and 91,000 US dollars, the first of them a 28 percent mobilisation invoice paid at kickoff to clean and prepare the data infrastructure and cloud before the first model runs, with each subsequent tranche released under a 96-hour client approval SLA and a phase-0 runway sitting ahead of phase 1. Along category, the same total resolves into seven budget lines, a 193,485 US dollar collaboration fee and 169,865 in salaries, scholarships and overhead together carrying more than two thirds of the contract, against 157,300 for all cloud and capital combined. We argue that these two structures are the actual risk-management instrument of the engagement: the mobilisation invoice funds the un-glamorous data plumbing that decides whether the science is even possible, the milestone gates convert an open-ended research bet into a sequence of small, cancellable ones, the SLA caps the schedule cost of the client's own review latency, and the people-heavy ledger names where the value and the risk really sit. The prior chapters of this programme, the manday-level pricing of a single phase and the year-two technology-transfer agreement, are pointers here rather than re-derivations; the subject is how the whole twenty months was funded.
Whitepaper
10 Mar 2026
21 pages · 20 min
Papering an AI Research Engagement: Effort Obligations, Phase-Capped Liability, and a Majority IP Split
An AI research engagement is usually written up through its model: the architecture, the ablations, the field metrics. The document that actually governs whether the work is buildable, survivable, and ownable is the one nobody publishes, the signed service agreement. We read that agreement from a three-phase subsurface-AI programme built for a major operator in Oman as a board and its legal counsel would read it, and we argue that its structure is a deliberate risk-transfer machine rather than boilerplate. Five clauses carry the transfer. The obligation is written as an effort, not a warranted result, because on genuinely uncertain research a results warranty is a payout the vendor cannot price. Liability is capped at the value of the phase in which a claim arises and never reaches a completed phase's fees, so exposure is a bounded staircase rather than the whole contract. Payment is due within fourteen days of the university receiving the end-client's funding, a back-to-back gate that keeps the vendor from financing the operator's own funding risk indefinitely while placing the vendor at the tail of the chain. Invoicing runs on a five-month cadence over a two-year term, and either party may exit on sixty days' notice, with the vendor able to leave cleanly after a completed phase. The foreground intellectual property splits fifty-one to the operator and forty-nine to the vendor and its company jointly, the smallest majority that still hands control to the operator while preserving the vendor's right to keep improving its own methods. We show the structure holding under load: when Phase-2 compute ran 2.17x over its budget, the contract's own change mechanism carried it as an additional request of USD 104,300 against a USD 533,000 approved programme rather than an emergency reopening. Three interactive instruments make the argument concrete. The deliverable is a reading of the paperwork that a technical buyer, a research vendor, and a board can each use to decide whether an AI engagement is structured to be signed, survived, and owned, or merely to be demonstrated.
Whitepaper
24 Feb 2026
18 min
Anatomy of a Fixed-Price AI R&D Phase: 250 Mandays, Five Roles, and the Discount Math
A fixed-price line item for an applied-AI research phase looks, to a buyer, like a single defensible number. It is not single and it is not a guess. This whitepaper reconstructs one such number from the engagement workbook: a 250-manday phase for a major operator in Oman, staffed by five roles on a fixed manday ladder (20 principal-QA, 60 senior researcher, 30 PM-plus-senior, 70 PhD researcher, 70 junior researcher), priced at academic day rates of 1600, 760, 760, 640 and 480 US dollars. Those academic rates run near half the open-market commercial ones (a principal at 1600 against 3200, a senior researcher at 760 against 1727), which is the single largest lever in the whole build and is available only because delivery routed through a university partnership. The role stack sums to 178,800 US dollars; a 35 percent bulk discount removes 62,580; a 7 percent contingency adds 8,135.40; and a 45,000 US dollar Phase-3 infrastructure package is carved out and priced separately, landing the phase at 169,355.40 US dollars. Behind the fixed price sits a two-party funding structure denominated in Omani rial: 165,000 OMR split 95,000 to the operator-facing firm and 70,000 to the university, each further split across a Phases 1-2 runway and the Phase-3 build. We argue that reading a fixed-price phase as an assembled artefact, rate arbitrage plus discount plus contingency plus carved-out compute, is the only way a buyer can tell a disciplined number from a hopeful one.
Whitepaper
24 Feb 2026
19 pages · 17 min
A Full-Stack Carbonate Subsurface-AI Reference Architecture: Fractures, Beddings, Vugs and Well-to-Well in One Pipeline
Carbonate reservoir characterisation has no single winning algorithm, and pretending otherwise is the most common way subsurface-AI programmes underdeliver. Fractures and bedding planes are sparse, geometric, supervised-learnable features best served by a set-prediction detection transformer; vugs are secondary-porosity blobs with no clean labels, best served by a tuned classical computer-vision pipeline; and field-scale continuity is a spatial-interpolation problem best served by geostatistics, not deep learning at all. A programme that masters one track and stops has automated a sliver of the interpretation and left the reservoir model incomplete. This whitepaper proposes a full-stack reference architecture that fuses all three, drawn from a roughly twenty-month, three-phase formation-evaluation engagement with a mid-sized Middle East carbonate operator we partnered with. We describe each track as an engineered component with its own architecture, training regime and accuracy envelope — a ResNet-10-backboned detection transformer reporting F1 of roughly 65% (fractures) and 63% (beddings) at a 3 cm depth tolerance and roughly 75% and 69% at 5 cm across fourteen wells; a multi-stage classical-CV vug quantifier validated on three wells against expert manual interpretation-software picks, running at under thirty seconds per metre versus four-to-five minutes for the prior morphological baseline; and a kriging-based well-to-well correlation layer that lifted interpretation productivity by roughly 60% and accuracy by roughly 75%. The contribution is not any single model. It is the routing-and-re-convergence pattern that carries one binary image log through three heterogeneous tracks and stitches their outputs into one reservoir-quality field model — and the argument that this fusion, not a better single detector, is what a carbonate operator actually needs to buy or build.
Whitepaper
11 Feb 2026
18 pages · 19 min
Contracting for Uncertainty: How to Paper an AI R&D Engagement
The standard commercial instinct for a paid engagement is to buy an outcome: define the deliverable, fix the price, and hold the vendor to a result. That instinct is exactly wrong for AI research and development, where the deliverable is a number nobody can commit to at signing. We reconstruct the contract we papered for a multi-year subsurface computer-vision programme with a major operator in Oman, and read every clause against the programme's own measured trajectory. Fracture-recall accuracy at a 5 cm depth tolerance climbed from roughly 10% at three labelled wells to 40% at eight to 75% at sixteen, a curve that was unknowable when the ink dried. A result guarantee written at kickoff would have committed the provider to a bar the science only cleared near the very end, and would have paid perverse incentives all the way there, rewarding safe, unfalsifiable claims over honest experiments. The architecture we used instead prices the work, not the outcome: an explicit effort obligation ('not responsible for achieving a specific result'), five tranches gated to phase boundaries, a data-timeliness clause that shifts delay cost onto whoever owes the data (which on this engagement was the client, whose labelled wells arrived five, then six, then eight of twenty-five across the year), provider-retained intellectual property with client indemnity on supplied data, a liability cap of USD 10,000 with all indirect damage excluded, invoicing every five months on 14-day terms, a two-year term, and provider phase-exit ramps after Phase 1 or Phase 2 on short notice. Three interactive instruments let a reader walk the argument: the map of the two contract modes against the science, the ledger of who absorbs a delay the provider did not cause, and the ramp that shows why bounded downside, not a performance promise, is what makes uncertain R&D signable at all.
Whitepaper
02 Dec 2025
20 pages · 19 min
After the Model Ships: Anatomy of a Year-Two AI Technology-Transfer Engagement
Most write-ups of an AI engagement stop when the model ships. The more revealing document is the one that comes next: the services agreement for the year after, which decides whether the operator ends up owning a capability or renting a dependency. We read the year-two agreement from a subsurface-AI programme with a mid-sized Middle East carbonate operator as a board would read it. The agreement funds USD 313,000 over a twelve-month term from 1 August 2023, payable either as twelve monthly instalments of USD 26,083.33 or four quarterly tranches of USD 78,250, and it resolves into exactly three pillars: AI technology transfer and training, delivered as six live sessions (three online, three onsite) archived on a learning-management system and climbing from AI essentials to hands-on work with the client's own transformer model; run-manage-operate for the on-premise Nvidia DGX A100 and the RADIX MLOps stack; and continuing AI engineering. The load-bearing structural fact is that the agreement is written on an effort obligation rather than a result obligation, with a 60-day termination notice, a 30-day cure, and a liability cap set at the phase amount. We argue that this is not incidental legal boilerplate but the mechanism of the handover: a vendor paid for effort, not for a held result, has no incentive to remain indispensable, and every one of the three pillars is therefore a transfer pillar, with the training archive, the run-book, and the engineering practice all designed to be absorbed by the client's own team. Transformation succeeds, on this reading, precisely when the vendor is contractually paid to make itself replaceable. The prior chapters of this programme, the R&D funnel and the phased operating model and the people trained, are pointers here rather than re-derivations; the subject is the contract that hands the whole thing over.
Whitepaper
18 Nov 2025
19 pages · 16 min read
Building Sovereign Subsurface-AI Capability: From an R&D Engagement to In-House Continuous AI
Most operators commission a subsurface-AI programme expecting to receive a model. What durably changes an asset is not the model — it is the capability to keep building models after the vendor leaves. This whitepaper reframes a three-phase, roughly twenty-month formation-evaluation engagement with a mid-sized Middle East carbonate operator we partnered with as a capability story rather than a modelling one. We trace the R&D innovation funnel that produced the algorithms — a NaN remover, unsupervised sinusoid pickers, DBSCAN and hierarchical clustering, a self-supervised segmenter, GAN and KNN imputation, and finally a detection-transformer lineage (DETR through a domain-specialised SUB-DETR) — and show how that funnel was industrialised into a productized Continuous-AI platform (automated fracture, vug, and well-to-well tools running on an on-premise MLOps control plane). Crucially, the programme trained 55 young professionals, 15 of them local nationals from regional universities, so the judgement to run, debug, and extend the platform stayed in-country. We argue that in-house capability is an industrialisation gate with its own acceptance test — the operator climbs the commercial phases themselves — and that a programme which builds the model but not the team has shipped a dependency, not a capability.
Whitepaper
30 Oct 2025
18 pages · 16 min
The Hardware Moat: Why Owning GPUs Beat Renting Them Through a Macro Shock
A multi-year subsurface AI programme makes a compute-sourcing decision on day one that looks like a finance question and turns out to be a risk question. Rent the GPUs and the programme's largest recurring cost is pinned to a market you do not control; own them and that cost is fixed the day you sign. We made the owning choice in 2020, before the 2022 energy shock, through a multi-year hardware partnership that put the cards on our own books and a local data-center deal that gave us full cost visibility. When the shock arrived, the consequences were measurable and asymmetric: the DGX and HPC hosting market climbed from USD 12-15K per month pre-war to past USD 20K post-war, energy prices rose about 400 percent, and electricity bills rose 394 percent, while the programme carried a single bounded, transparent contingency of OMR 20K, roughly USD 51K, across a twelve-month window. This paper reconstructs the decision as a board would: it reads the three-way cost model behind the choice (public cloud near EUR 17,200 a month, private cloud near EUR 4,000, on-premise near EUR 20,000, with a mid-size AI server at 50-120K and two servers at 100-200K amortised), it shows why the public-cloud and on-premise sticker prices land alike while their risk profiles diverge, and it frames the contingency as the single approvable number that ownership put on the agenda instead of an open-ended rental bill. We argue that owned compute is best evaluated as a hedge with a known premium, not as a line of capital spending, and we end with the concrete questions a CIO should ask before committing to a multi-year AI programme's infrastructure.
Whitepaper
28 Oct 2025
20 pages · 20 min
After the R&D Ends: Three Ownership Models for the AI You Just Built
When a multi-year subsurface-AI R&D engagement ends, the operator is holding a working model and a decision most programmes defer until it is urgent: who owns the compute that will train the next version, and the one after that. The model is portable; the machine that makes new models is not, and the choice of who runs that machine sets the operator's cost base, its staffing, and how fast it can respond to new wells for years. This whitepaper generalises the transition menu we put in front of a mid-sized Middle East carbonate operator at that gate into three ownership models plotted on one curve. Tier one, replicate the stack in-house, keeps full control but requires the operator's own ICT to grow AI muscle and runs a full-model retrain in two-to-three weeks on hardware the operator buys and staffs. Tier two, a light or medium managed service on the operator's premises, lowers cost and internal lift and stretches a full retrain to days. Tier three, as-is managed supercompute, is zero lift-and-shift because the R&D engagement already paid the upfront investment, so a full retrain lands in minutes-to-hours for a fixed annual fee. We argue three things with three instruments. First, that per-model training time and internal-capability burden move in opposite directions across the tiers, which is why the decision is a curve and not a ranking. Second, that the train-time collapse is bought by climbing three orders of magnitude of compute, from a workstation GPU stack through a DGX A100 node to a multi-node SuperPod, and that only the fixed-fee tier puts the top rung under a retrain without a capital purchase. Third, that which tier fits is decided by three concrete questions about the operator's ICT rather than by the fastest train-time on offer. The deliverable is a way to price the R&D-to-industrialization gate as the ownership decision it is, before the last progress report is signed.
Whitepaper
23 Sept 2025
18 pages · 16 min read
From Manual Picks to a Continuous Interpretation Workflow: Operationalizing AutoFrac and AutoVug for Reservoir Engineering
Borehole-image interpretation is run almost everywhere as a batch craft: a senior geoscientist takes a single high-resolution borehole image log, picks bedding planes, fractures and vugs sinusoid by sinusoid over a multi-week pass, hands a static result to the asset team, and moves to the next well while the queue grows. This whitepaper argues for replacing that batch craft with a continuous interpretation workflow, drawn from a roughly twenty-month, three-phase formation-evaluation engagement with a mid-sized Middle East carbonate operator we partnered with. We describe the two production capabilities at its core — AutoFrac for bedding-and-fracture detection and AutoVug for vug analysis, each running roughly five times faster than manual picking — not as models but as stages in a standing pipeline a well moves through: ingest, normalise, patch, infer, validate, and emit a structured eight-column output of predicted bedding planes, sinusoids, fractures and vugs that becomes the interface to reservoir engineering. We treat blind-zone validation on a held-out well and depth interval as the pipeline's standing acceptance gate rather than a one-time benchmark, and we show how the interpreter capacity reclaimed by the speed-up is redeployed — onto the operator's 80-plus-well backlog, and into continuous vug-percent and fracture-density reservoir maps that a manual team could never have produced at scale. The reframing is the point: interpretation stops being a three-to-four-week-per-well event and becomes a running service that reservoir engineering can query.
Whitepaper
24 Jun 2025
19 pages · 20 min
Compute Budgeting for Vision Training Under a Hard GPU-Memory Ceiling
Conventional training-cost planning treats compute as an arithmetic problem: count the floating-point operations, divide by device throughput, multiply by a price. That arithmetic quietly assumes a fixed-shape batch tensor, and for a raster well-log archive that assumption never holds. The images span roughly 3,200 to 12,800 pixels wide against a narrow 480 to 640 pixel height, in no fixed aspect ratio, so they cannot be stacked, and the first cost that bites is not arithmetic at all; it is the GPU memory committed by the single widest member of a mini-batch. This whitepaper is the first-person engineering and financial account of training VeerNet, our encoder-decoder segmentation network with a transformer attention bottleneck, under exactly that ceiling. The binary segmentation stage was forced to batch size 1, which trained correctly and finished 50 epochs over 2,000 instances in 110 minutes, but starved the optimiser of gradient averaging. The three-class multiclass stage on 15,000 synthetic instances reached an effective batch of 16 through a custom collate function that pads each mini-batch to its widest member on a 32-pixel alignment grid and a validity mask that keeps padded pixels out of the loss; it finished 50 epochs in 550 minutes. We set out the memory inequality that governs the whole problem, the alignment-padding cost that the encoder stride structure imposes, and the two rentable GPU tiers, 750 and 1,800 EUR per month, that the engagement budgeted against. We then build the spend plan an executive can defend in a budget review: how to scope rented hours to a batch-size-1 wall clock, why the widest log, not the average, sizes every line item, and which engineering levers move the bill versus which only move the FLOPs. The deliverable is a method for budgeting compute when the binding constraint is memory, and a worked ledger that an upstream operator or any team training on variable-dimension imagery can reuse.
Whitepaper
17 Jun 2025
19 pages · 17 min read
Deploying Subsurface AI Inside a Closed NOC Network: The On-Prem MLOps and Geo-Platform Integration Blueprint
Most subsurface AI fails its final exam not on accuracy but on address: the model works, but it works in a cloud the operator's security team will never sign off on. This whitepaper is the deployment blueprint for the other path — productionizing a deep-learning formation-evaluation system entirely inside an oil-and-gas operator's air-gapped corporate network, where no training byte, no checkpoint, and no log image is permitted to leave the perimeter. Drawn from a multi-phase engagement with a mid-sized Middle East carbonate operator we partnered with, it specifies the four layers that make on-prem subsurface AI real: an on-premise GPU compute tier (an Nvidia DGX A100 class server backed by a data-management server and custom GPU nodes, replacing the consumer-grade cards the R&D phase ran on); an MLOps control plane built on a custom platform with Weights & Biases experiment tracking, a Seafile data backbone, GitHub Enterprise and Ubuntu; a delivery surface that ships each model as a containerized application exposed over a LAN-only API into the operator's existing geomatics platform; and a confidentiality regime — including depth-digit masking — that lets the system retrain and serve without exposing the operator's subsurface position. We then make the decision explicit: three hosting models trade operator control against operating burden, and they are not equivalent on the one axis that matters in production — retraining latency, which ranges from minutes-to-hours under a managed on-prem arrangement to days when the operator runs the loop alone to two-to-three weeks off-premise. The handover is the acceptance test; this whitepaper is the wiring diagram behind it.
Whitepaper
11 Jun 2025
18 pages · 18 min
Anatomy of a Top-K Mode-Subtraction Vug Detector: Five Enhanced Copies of One Log
Automated vug quantification on carbonate borehole-image logs has a published body of work on each of its stages in isolation: mode subtraction to suppress the matrix, adaptive thresholding to resolve pores, a circularity gate to reject fractures, a false-positive filter to clean the catalog. What a reviewer asks for is different: not the best stage, but the whole pipeline stated precisely enough to reproduce and defend. This whitepaper is that statement. One borehole-image patch is min-max scaled to 0-255, then turned into five mode-subtracted copies (top-K modes K=1 to 5) at a minimum intensity spacing of delta-m 5 so the copies are genuinely distinct. Each copy is adaptively thresholded with a 31x31 Gaussian-modulated block, a window that covers 46.58 cm2 of borehole wall, at a one-pixel step. Contours are extracted from every copy and each candidate is validated with a Laplacian-variance and mean-difference test on a bounding box expanded 10 percent per side, with a mean-based enlarged radius of 2R and cross-copy duplicates merged at an intersection-over-union of 0.2. The output is an interval signature: per-vug area, predominantly 1 to 6 cm2; circularity, distributed 0.28 to 0.85 with a semi-circular peak in 0.45 to 0.7; and azimuth, resolved into four quadrants. We present the circularity ridgeline spectrum and the estimated-vs-manual vug track as the two instruments that make the case, and we keep to the constants that survived review: k=5, delta-m 5, block 31, IoU 0.2, 10-percent expansion, runtime near 2 minutes per 2 m interval with 100 threshold values for path opening. The argument is that the pipeline's defensibility comes from the validation stage, not the detection stage.
Whitepaper
28 May 2025
18 pages · 19 min
One Classical-CV Parameter Set, Every Well: A Field Guide to the Eight Knobs of the Vug Detector
A classical computer-vision vug detector for carbonate borehole-image logs is attractive precisely because every decision is a named, tunable threshold rather than a weight in a checkpoint. That legibility only pays off if the parameters are portable: an operator retraining on a new well needs to know exactly which numbers travel unchanged and which must be re-established, and on what evidence. This whitepaper states the contract in full. Eight interpretable knobs govern the pipeline: top-k intensity modes (k=5) with a minimum mode separation (delta-m=5), a centroid-difference gate (tau=5 pixels), an intersection-over-union merge threshold (0.2), a Laplacian bounding-box expansion (10% each side), a mean-based enlarged radius (2R), a block size (31) and an adaptive-threshold constant (C=mean patch intensity). Six of these are universal, fixed once for every well, orientation and imaging tool; block size and C are recommended starting points; and exactly two, the Laplacian smoothness threshold and the mean-based darkness threshold, need a one-time per-well base value calibrated from one or two representative sections and then frozen for the whole well despite grayscale drift with depth. We also document two corrections that peer review forced and that materially improved the detector: a sign check on the mean difference that stops opposite-spectrum minerals being labelled vugs, and a change from deleting one of two overlapping contours to merging them, which preserves the true vug extent. The result is a rare thing in applied geoscience CV: a fully specified, reproducible pipeline whose portability is written down rather than assumed.
Whitepaper
19 May 2025
20 pages · 22 min
Recovering Trapped Data: A Five-Year View of Industrial Digitisation
Look back across five years of raster-to-vector work and the individual projects stop looking like separate projects. The same shape recurs every time: a huge archive of scanned documents nobody can query, a synthetic-data bootstrap that manufactures the training set the archive never had, a learned segmenter that turns paper into trustworthy tabular data, and a marginal cost per recovered document that keeps falling because one trained model amortises across the whole corpus. This whitepaper reads that path at portfolio altitude, using the sourced numbers from a reference engagement, and argues that the repeatability is the asset, not any one model.
Whitepaper
09 Apr 2025
18 pages · 17 min
Vug Fingerprints, Not Vug Percentages: Per-Interval Morphology as a Reservoir-Quality Log
Secondary porosity in a carbonate reservoir is usually reported as a single vug percentage per interval, a scalar that two very different rocks can share. We describe a different deliverable, built and validated with a mid-sized Middle East carbonate operator: a continuous per-interval morphology log. Every 2 m of borehole yields a six-track statistical fingerprint - a kernel-density estimate of vug area (observed 1 to 12 square centimetres, mass concentrated between 1 and 3.5 and predominantly 1 to 6), a kernel-density estimate of circularity (0.28 to 0.85, peaking semi-circular between 0.45 and 0.7), a four-quadrant azimuth spectrum (0 to 90, 90 to 180, 180 to 270, 270 to 360 degrees), a vug count, the vug percentage itself defined as vug area over interval area, and an estimated-versus-manual overlay track that validates the automated estimate interval by interval. Stacked down the well, these tracks are a log in the petrophysical sense: you read them vertically, you spot the interval where the character changes, and you correlate that signature to the same signature in the next well. The percentage is one of the six tracks, not a summary of the other five. We show that two intervals carrying an identical percentage separate cleanly in the area and circularity tracks and in the azimuth spectrum, that reservoir-quality populations occupy distinct regions of the area-circularity plane a scalar cannot see, and that well-to-well correlation - a Phase-3 objective of the programme - is tractable on the multi-track signature and ambiguous on the percentage alone, because many intervals repeat the same percentage. The statistical outputs are computed on a fine grid (per 10 cm: total count, total area, mean area, standard deviation of area, and the area, circularity and azimuth spectra), and the whole log is produced without a human picking individual pores. The argument is not that percentages are wrong. It is that a percentage is a projection, the fingerprint is the data, and a reservoir-quality workflow should consume the data.
Whitepaper
08 Apr 2025
16 pages · 16 min read
From 3-4 Weeks to 2 Hours per Well: The Economic Case for Automating Borehole-Image Interpretation
Borehole-image interpretation is the most expert-bound, least scalable step in formation evaluation. A senior geoscientist spends days to weeks picking bedding planes, fractures and vugs off a single high-resolution borehole image log by hand, and the queue of unanalysed wells only grows. This whitepaper makes the economic case for automating that step with a productized deep-learning pipeline, drawn from a roughly twenty-month, three-phase engagement with a mid-sized Middle East carbonate operator we partnered with — a programme scoped against 80-plus processed image logs. We separate three returns that are usually conflated: the cycle-time compression of a single well from a multi-week manual pass to a roughly two-hour reviewed, agentic pass; the productivity and accuracy dividend the operator's interpreters captured once the system was productized — well-to-well correlation lifting interpretation productivity by about 60% and accuracy by about 75%, with the fracture and vug detectors running roughly five times faster than manual picking; and the strategic return of closing the gap between the handful of wells a manual team can interpret and the full multi-well backlog the asset actually holds. We then show why these returns only persist if the system is engineered to be owned by the operator, and we put a frame around what the reclaimed capacity is worth when it is redeployed rather than simply banked. The argument is deliberately commercial: the model is a means; the interpretation throughput it unlocks is the asset.
Whitepaper
11 Mar 2025
16 pages · 16 min read
Data-Centric AI Under Extreme Well Scarcity: How 235 Patches Became 75,000 and Why 14 Wells Beat Synthetic Overload
In most proprietary subsurface domains the binding constraint is not compute or architecture — it is wells. You cannot acquire more of them, you cannot buy a public corpus that matches your formation, and the labels you do have were written by a handful of expert interpreters. This whitepaper is a field account of building a borehole-image detector under exactly that constraint: fourteen wells of high-resolution borehole image logs from two different microresistivity imaging tools, from a mid-sized Middle East carbonate operator we partnered with, and a reservoir interval that started life with thirty-two labelled sinusoids. We show how the model quality that mattered was won data-side rather than model-side. Overlapping-window patching plus a six-transform augmentation regimen applied ten times per labelled patch scaled the working sinusoid set roughly sixty-five-fold — from about nine hundred image-and-ground-truth pairs to over fifty-five thousand, and from one interval's two hundred thirty-six patches to four thousand two hundred twelve. Augmentation alone collapsed a classification error from a degenerate one hundred percent to 2.618 percent. Adding real wells, three to fourteen, drove classification error from 93.115 to 2.536 — and crucially, it beat the obvious shortcut of inflating the corpus to ninety-two thousand augmented samples, which made the model worse. The lesson generalises: in scarce proprietary domains, data engineering is the model engineering, and not all data is the kind that helps.
Whitepaper
18 Feb 2025
21 pages · 22 min
From One-Off to On-Demand: Productising a Raster-Log Digitiser into a Hosted Scan-to-LAS Service
A research model that wins on a validation set is not a product, and the distance between the two is mostly engineering and product design rather than more training. We built VeerNet, an encoder-decoder segmentation network with transformer attention refinement, as a bespoke deliverable for a US onshore operator: it digitised scanned raster well logs into spec-compliant LAS and CSV files, trained on roughly 15,000 synthetic curves, reaching a peak coefficient of determination of 0.9891, a lowest mean absolute error of 0.0132, and a 20x to 200x speedup over manual tracing. This whitepaper documents the productisation of that one-off capability into a repeatable, hosted, self-serve service, where a user uploads a scan, the model digitises it, a reviewer corrects the imperfect prediction, and a compliant file is downloaded. We treat productisation as three coupled engineering decisions. First, the shape change: a bespoke project is built once and billed once, while a service amortises one trained model across many scans, which inverts the unit economics and forces a different posture toward cost, isolation, and idempotency. Second, the flow: the four self-serve stages of upload, digitise, review, and download, where the review stage is not a convenience but the load-bearing element, because a peak curve-class intersection-over-union of only 0.51 means the model is excellent at most of the trace and wrong on a minority of it, and the product must let a human see and fix exactly that minority. Third, the throughput dividend: one build pointed at a reference archive of 136,771 scanned images returns tracing hours that scale with the speedup, the compounding payoff that justifies the productisation work. We embed three interactive instruments, a bespoke-versus-hosted ledger, the upload-to-download product flow, and a one-model-many-scans throughput meter, and close with the working rule we adopted, that the model is the easy half and the service is the half that decides whether anyone outside the original engagement ever benefits from it.
Whitepaper
11 Feb 2025
17 pages · 15 min read
MLOps for Subsurface AI: Reproducibility, Versioning and Handover
Most subsurface AI programmes are run as research and die as research. They produce a notebook that beats the baseline, a deck that impresses a steering committee, and a model nobody can rebuild a year later. This whitepaper is a field guide to the alternative: running a multi-year subsurface ML programme as engineering. Drawn from a three-phase, roughly twenty-month formation-evaluation engagement with a mid-sized Middle East carbonate operator, it covers the four disciplines that decide whether a trained system survives — data versioning with DVC, experiment tracking with Weights & Biases, CI/CD and serving, and bit-level reproducibility through seeding and frozen checkpoints — and then the discipline almost every programme skips: the handover that puts the model, the pipeline, the runbooks and the know-how inside the operator's in-house ICT team. We argue that the handover is not a closing courtesy. It is the acceptance test for everything built before it, and a programme that cannot be handed over was never engineered in the first place.
Whitepaper
09 Dec 2024
20 pages · 21 min
Half the Bits, Twice the Headroom: Mixed-Precision Training for 12800-Pixel Log Images
A training step has one cost that scales with the size of the input and dominates everything else on an encoder-decoder with skip connections: the activation memory held alive from the forward pass for the backward pass to consume. For a raster well-log archive whose widest image is 12,800 pixels across by up to 640 tall, that single term is what breaches the GPU memory ceiling, and it does so at batch size 1, where the usual relief valves of smaller batches and gradient accumulation have nothing left to give. This whitepaper is the first-person methodology account of recovering the headroom from the numerics rather than from the batch size or the wall clock. The central move is to store activations in 16 bits instead of 32 under automatic mixed-precision autocast, which halves the dominant memory term at 2 bytes per value against 4, while keeping the loss, the normalisation statistics, and the master weights in fp32 so accuracy does not move. We set out the activation-memory inequality the whole problem is governed by, the two half-precision formats the loop can choose between, and the reason the choice between them is a numerics decision and not a convenience: bfloat16 keeps the full fp32 exponent range with 8 exponent bits and needs no gradient rescue, while fp16 keeps only 5 exponent bits, underflows small gradients to zero, and must be paired with a dynamic loss scaler that multiplies the loss by a power of two before the backward pass and divides it out before the optimiser step. We walk three instruments that argue the budget, the underflow mechanism, and the format anatomy, then state the operating posture we adopted for the 12,800-pixel worst case at batch size 1, the GroupNorm group-size floor of 16 that keeps normalisation valid at that tiny physical batch, and the failure modes that send a half-precision run divergent. The deliverable is a method for buying back activation headroom with numerics that any team training on variable-dimension imagery under a hard memory ceiling can reuse.
Whitepaper
19 Nov 2024
18 pages · 18 min read
Well-to-Well Domain Adaptation: Carrying a Fracture and Bedding Model Across Wells with Kriging and Stratigraphic Correlation
A borehole-image detector trained on one set of wells is, by default, a single-well predictor. It reads the wall of the borehole it was shown and nothing beyond it, and when a new well arrives from an adjacent part of the same field its accuracy can fall off a cliff that has nothing to do with the model and everything to do with distribution shift. This whitepaper is a field account of closing that gap. Working with a Middle East NOC / carbonate operator we partnered with, across a multi-phase formation-evaluation engagement, we took a Detection-Transformer-derived fracture and bedding detector and made it travel — not by re-labelling every new well, but by treating well-to-well transfer as a domain-adaptation problem solved with geostatistics. The detector emits two engineered logs, a fracture-density curve and a bedding-density curve, computed at ten-centimetre resolution; 2D and 3D kriging with gaussian, linear and exponential variograms interpolate those logs between wells spaced roughly forty to eighty metres apart; and stratigraphic well-tops align the interpolation within a common formation rather than across geological apples and oranges. The result is a field-scale correlation engine: bedding density that correlates laterally between neighbouring wells, a productivity lift of roughly sixty percent and an interpretation-accuracy lift of roughly seventy-five percent against a ninety-five-percent target precision and ninety-percent stratigraphic-correlation success. We argue that the unit of subsurface AI is not the well. It is the field — and the engineering that gets you there is feature-space alignment, not a bigger model.
Whitepaper
13 Nov 2024
21 pages · 22 min
A Phased Operating Model for Standing Up Subsurface AI Capability
Most operators who decide to build subsurface AI in-house frame the decision as a procurement question, which model, which vendor, which platform, and then discover that the model was the easy part. What they were actually deciding was whether to stand up a capability, and a capability is not a deliverable; it is a governed sequence of phases that each retire a specific class of risk before the next begins. This whitepaper sets out the operating model we run for that build-out: four phases, assess, pilot, scale, and operate, mapped onto the four project blocks the engagement is scoped in, each with its own roster of roles, its own exit gate, and its own staffing posture. We argue three things with three instruments. First, that the progression is genuinely governed and not merely sequential: each phase has a distinct staffing shape and a concrete exit test, the pilot against a curve-reconstruction fidelity gate at peak R-squared 0.9891, the operate phase against a monthly compute run-rate of 750 to 1800 EUR, and a gate skipped to save time does not save the risk it was meant to retire but forwards it downstream where it costs more to fix. Second, that the choice between an accelerated sixteen-week track at six full-time staff and one hundred eighty thousand EUR and a standard thirty-two-week track at four staff and one hundred thousand EUR is a concurrency-and-premium decision rather than a scope decision: the fast track does the same work sooner with more people for more money, not less work. Third, that the operate phase is a low standing run-rate whose floor is set by the served model's compute and is unmoved by how well the earlier phases were run, so the temptation to shortcut assess and pilot buys nothing at the stage where the programme actually lives. The deliverable is a way to plan, staff, and govern an in-house subsurface-AI build-out as the phased capability programme it is, rather than as a model handover with a support contract stapled to it.
Whitepaper
27 Aug 2024
18 pages · 16 min
Beyond Bulk Percentages: AI-Assisted Individual Vug Quantification in Carbonate Reservoirs
Secondary porosity in carbonate reservoirs is dominated by vugs, and the way it is reported in practice — a bulk vug percentage per interval, hand-picked on a borehole image log — discards almost all of the information that controls reservoir quality. A 6 percent interval made of a few large connected vugs and a 6 percent interval made of many small isolated ones flow very differently, yet the bulk percentage cannot tell them apart. This whitepaper presents a classical computer-vision pipeline — deliberately not deep learning — that quantifies vugs individually rather than in bulk. The pipeline chains top-K mode subtraction for local-variation enhancement, Gaussian-modulated adaptive thresholding, Suzuki-Abe contour extraction, an area-and-circularity geometric gate, centroid-and-IoU aggregation to de-duplicate, and a two-stage false-positive filter (Laplacian-variance contrast and mean-intensity differencing) that suppresses the conductive bedding planes and fractures a naive thresholder would mistake for vugs. The output is not a percentage but a distribution: per-vug area (observed 1 to 12 square centimetres, concentrated between 1 and 3.5), circularity (0.28 to 0.85, peaking semi-circular at 0.45 to 0.7), and an azimuth spectrum, computed every 0.1 metre. Validated on three wells spanning both vertical and horizontal high-resolution borehole image logs acquired with two different microresistivity imaging tools, using a single shared parameter set, the pipeline matches the expert's interpretation-software ground truth and, at several intervals, recovers dissolution vugs the manual interpreter missed. Because every step is an interpretable operator rather than a black box, the result is a defensible, auditable, per-vug catalogue — and a natural source of ground-truth masks for a future supervised vug-segmentation model. We frame the work as applied computer-vision engineering: the geoscience defines the gate, the algorithm enforces it reproducibly, and the production layer makes it run on two-hundred-metre logs without a human in the loop.
Whitepaper
27 Aug 2024
20 pages · 22 min
Unit Economics of a Curve-Digitisation SaaS for Subsurface Teams
There is a well-worn failure mode in applied machine learning: a team builds a genuinely good model, digitises a hard corpus that nobody else can, and then discovers the economics of selling it were never worked out. This whitepaper is a product-and-go-to-market reading of a vertical curve-digitisation SaaS for subsurface teams, written from the position of having built the digitiser and then having to price it. It is deliberately not a re-telling of the model architecture, which we have published elsewhere; it is about the numbers a business runs on. The argument has one spine. A vertical SaaS becomes attractive only when the price a seat pays sits far above the marginal cost of serving that seat, and only when a disciplined market funnel keeps the revenue target honest rather than a number a slide deck asserts. We ground it in the sourced figures from one reference programme: a seat priced at 1,200 USD per user per year, a marginal serving cost derived from a GPU that rents for 750 to 1,800 EUR per month, a total market measured at 134B USD, a serviceable slice of 6.7B USD, a 3 percent obtainable share by end of year five reaching a 180M USD revenue projection against 6M USD in the first twelve months, a 3M USD seed with an 18-month runway, and two build tracks, an accelerated one at 180,000 EUR and a standard one at 100,000 EUR. Three instruments carry the argument: the contribution margin per seat, the seed as a fixed monthly burn ceiling, and the delivery-track fork where speed to revenue repays its own premium. The point is that the research being good is a precondition, not the business case, and the business case has to be built on numbers that survive division.
Whitepaper
18 Jun 2024
22 pages · 22 min
AI for Long-Term Energy Project Engagement and Stakeholder Management
Long-lived subsurface projects, and carbon capture and storage most of all, carry two machine-learning problems that people keep collapsing into one. The first is site selection, a bounded vision task: read the operator's legacy raster archive, of the order of 136,771 scanned log images, digitise the curves that constrain the reservoir, and choose where to store. A pure convolutional digitiser does this and is graded on the reconstructed curve, where our own work peaks at R-squared 0.9891. The second problem begins the day the site is chosen and runs for the life of the project: the community relationship, the regulatory posture, and the stakeholder engagement that decide whether a storage site that is geologically sound is also socially durable. That second problem is not a segmentation task with a different head; it is a different model family. This whitepaper sets out the case for a hybrid fuzzy-logic-plus-deep-net architecture as that family, anchored on the published result that such a model forecasts CO2-storage project engagement at 90.476% accuracy, and argues three things with three instruments: that the fuzzy-DNN engagement forecast reaches usable accuracy as a distinct family from the digitiser, that the two families diverge on every axis of the problem that matters, and that the value of the forecast is not that it freezes engagement but that it buys the lead time to defend a decaying-trust asset before it crosses the acceptance floor. The deliverable is a way for an operator to organise its subsurface AI portfolio so the model that reads the rock and the model that reads the room are built, graded, and governed as the two different things they are.
Whitepaper
14 May 2024
19 pages · 17 min
GeoBFDT: End-to-End Detection Transformers for Fracture and Bedding Picking in Carbonate Image Logs
Automated fracture and bedding picking on borehole image logs has been dominated by mask-based detectors that find one geological feature per image and lean on hand-tuned post-processing — a poor fit for fractured carbonates, where sinusoids overlap, intersect, and crowd a single patch. We present GeoBFDT, an end-to-end Detection Transformer that reframes the task as set prediction: a lightweight ResNet-10 backbone trained from scratch feeds a four-layer transformer encoder and four-layer decoder, a fixed bank of learned object queries emerges as candidate detections, and a Hungarian bipartite-matching loss assigns each query to at most one ground-truth sinusoid before any gradient is computed. The model regresses depth, dip and azimuth directly from the image — no masks, no anchors, no non-maximum suppression. Trained on 14 vertical wells (two different microresistivity imaging tools) from a fractured Middle East carbonate play we partnered on, GeoBFDT reaches an F1 of roughly 65% for fractures and 63% for beddings at a 3 cm depth tolerance, rising to roughly 75% and 69% at 5 cm, with dip accuracy near 90% at 3 degrees and azimuth accuracy near 92% at 15 degrees. Ablations isolate the choices that matter: a ResNet-10 backbone outperforms ResNet-18 and ResNet-34 by an order of magnitude on classification error on this small-data regime, sinusoid-aware augmentation moves classification error from 100% to 2.6%, and dynamic-image input beats static. We argue that set prediction — not bigger masks — is the correct inductive bias for multi-fracture carbonate image logs, and that the engineering, not the geoscience, is what makes it production-grade.
Whitepaper
18 Mar 2024
18 pages · 17 min
Evaluating Curve-Extraction Quality Beyond IoU and F1
Semantic-segmentation models are conventionally graded on pixel-overlap metrics, Intersection-over-Union and F1, and a raster-log digitiser built as a segmentation network inherits that habit. We argue the habit is wrong for this product class. The thing a digitiser delivers is not a pixel mask; it is a one-dimensional curve, a value at every depth, exported as a CSV that a petrophysics package consumes. When VeerNet, our encoder-decoder segmentation network for raster well-log digitisation, is graded on the mask it scores a peak IoU of 0.51 and an F1 of 0.55, with per-class IoU of 0.94 on background but only 0.26 and 0.21 on the two curve classes that actually carry the signal. Read in isolation those numbers look like a model that is not ready. Yet the same checkpoints, evaluated on the reconstructed curves they produce, reach a peak coefficient of determination of 0.9891 against the native LAS data, a lowest mean absolute error of 0.0132, and a lowest mean squared error of 0.0004. The gap between those two readings is not noise; it is structural, and it follows directly from what a one-pixel-wide trace does to an overlap metric. This whitepaper explains why pixel-space metrics systematically undersell thin-structure segmentation, defines a deliverable-space evaluation protocol built on per-curve R-squared, MAE, and MSE measured against ground-truth LAS at fixed depth sampling, walks the per-loss numbers from our five-loss ablation under both readings, and states the operating rule we now apply: a digitiser is graded on the curve it emits and the human-verification step that follows, not on the mask it passes through on the way there.
Whitepaper
20 Nov 2023
19 pages · 20 min
An MLOps Blueprint for Vision Models in Resource-Constrained Industry
The published MLOps playbook is written for teams that can scale out of any problem: add GPUs when memory is tight, add engineers when the schedule slips, add budget when the cloud bill climbs. Industrial computer-vision work rarely has those exits. We built VeerNet, our encoder-decoder segmentation network for digitising scanned raster well logs into LAS-grade curves, for an onshore Texas operator under three simultaneous hard ceilings: a GPU memory limit that forced the binary segmentation stage to a physical batch of one image per pass, a headcount cap, and a fixed price. This whitepaper is the blueprint we ran against, written from the inside. We treat the memory ceiling as the constraint that sets every other decision, and show how a custom collate function that pads each mini-batch to its widest member and masks the padding out of the loss recovered an effective batch of sixteen on the harder three-class multiclass stage. We cost training in the only unit that was actually scarce, wall-clock time, anchored on the 110 minutes the binary stage took for fifty epochs over two thousand images and the 550 minutes the multiclass stage took for fifty epochs over fifteen thousand. And we tie a deliberately compact architecture, a five-block encoder, a five-stage decoder, and two transformer attention layers on the bottleneck, to a delivery envelope of sixteen weeks, six full-time engineers, and 180,000 EUR, against a standard track of thirty-two weeks, four engineers, and 100,000 EUR. The argument throughout is that constraints are not obstacles to the blueprint; they are the blueprint. Three interactive instruments let a planner walk the memory ceiling, the training-time budget, and the delivery envelope directly.
Whitepaper
14 Aug 2023
19 pages · 20 min
Missing-Data Diagnostics Before Petrophysical Machine Learning
Petrophysical machine learning treats missing data as a preprocessing afterthought, a line of code that drops or fills the gaps so training can begin. We argue the opposite: that the most consequential modelling decisions are taken before any model exists, in a diagnostic pass that profiles, visualises, and reasons about missingness, and that skipping that pass is the most common and most expensive mistake in subsurface ML. A well-log table is missing data in three distinct registers at once. At the column level, the public Norwegian-Sea teaching slice (118 wells, 22 electrical-measurement columns, profiled with the missingno library by McDonald) shows a handful of near-complete primary curves above a long tail of sparse specialist tools, and the keep-or-drop decision on each column changes what the model can even learn. At the corpus level, a scanned regulatory archive is missing labels wholesale: the public Texas Railroad Commission collection holds 136,771 raster TIF files against only 7,781 digital LAS files, so at most a few percent of the imagery can ever be paired into supervised examples, which is the entire reason a 20,000-log synthetic corpus had to be manufactured. At the pipeline level, a missingness decision taken or skipped upstream propagates silently through the 80/20 train/validation split to the validation metric, where our multiclass model recorded a background IoU of 0.94 against a curve-1 mask IoU of just 0.26. We set out a diagnostic protocol with three steps, profile then visualise then decide, walk three interactive exhibits that argue each register of the problem, and state the operating rule we now apply: no corpus enters training until its nullity has been profiled, its label coverage has been mapped, and the consequences of every drop, fill, or synthesis have been reasoned about in the open. This document is distinct from our companion whitepaper on curve-extraction evaluation; that one grades the deliverable after the model runs, this one disciplines the data before it does.
Whitepaper
12 Jun 2023
19 pages · 18 min
Encoder-Decoder Backbones for Variable-Aspect-Ratio Log Images
Raster well-log digitisation has to cope with images whose width varies by a factor of four, from a 3,200 by 480 floor to a 12,800 by 640 ceiling, with no fixed aspect ratio in between. The conventional response is to make the input regular before the network sees it, by resizing or by tiling into fixed crops, and then to spend modelling effort on the loss function. We argue the priority is inverted. CurveNet, the encoder-decoder backbone we built for this task, accepts the full variable-aspect-ratio range natively because its shape, not its objective, is built to absorb size: five stride-2 encoder stages divide the spatial dimensions away while a 128-dimensional bottleneck holds constant, two transformer attention layers operate on that fixed-width bottleneck regardless of input length, a GroupNorm grouping floored at sixteen keeps normalisation statistics stable when the batch is small, and five upsample decoder stages reconstruct a mask at the original resolution. The same architecture trains a sixteen-image multiclass batch through a custom collate path and a single-image binary batch under a tighter memory budget without any structural change. This document makes the case that this shape is the load-bearing decision. It walks the down-up cadence with a draggable stress bench that stretches the input across the full range, examines the attention bottleneck that gives the network a global receptive field over a long thin image, and traces how the variable-dimension batch is fed without tiling. It situates the design against the public encoder-decoder, normalisation, and attention literature it builds on, reports the figures the architecture made reachable, and states the engineering rule we now follow: choose the backbone for the input distribution first, and treat the loss as a tuning knob on top of a shape that already fits.
Whitepaper
22 May 2023
20 pages · 21 min
Sub-Pixel Centrelines and Depth Registration: A Validation Framework for Digitised Curves
A semantic-segmentation digitiser of scanned well logs is conventionally accepted on a mask-overlap criterion, but the artefact it actually delivers is a one-dimensional curve of values indexed by true measured depth, and mask Intersection-over-Union is blind to the two error sources that decide whether that curve is correct. The first is centreline localisation: a segmentation head emits a probability band several pixels wide at every depth row, and reducing it to a single position by integer per-column argmax snaps the recovered curve to the pixel lattice and caps its accuracy at half a pixel regardless of mask quality. The second is depth registration: the pixel rows of a scan carry depth only through a header scale, and an affine pixel-to-depth map fitted from the header anchors propagates any header mis-read into a depth-dependent value error before any curve interpolation runs. This whitepaper presents the validation framework we built for VeerNet, our encoder-decoder segmentation network for raster well-log digitisation, in which centreline accuracy and depth alignment, not mask IoU, are the acceptance criteria. We define a parabolic three-point sub-pixel refinement of the per-column argmax that recovers fractional position between grid lines, the two-anchor affine depth-registration map and its residual structure, and an end-to-end protocol that grades the reconstructed curve against ground-truth values resampled to a common axis at 300 interpolated depth points using per-curve R-squared, mean absolute error and mean squared error. On that protocol VeerNet reaches a peak coefficient of determination of 0.9891, a lowest mean absolute error of 0.0132 and a lowest mean squared error of 0.0004, even though the same checkpoints score a peak mask IoU of only 0.51. The framework ships three interactive exhibits, a sub-pixel centreline refiner, a depth-registration ladder, and an end-to-end residual scanner, that let a reviewer move each error source independently and read its consequence on the deliverable. We close with the methods needed to reproduce the protocol, its limitations, and the acceptance rule we now apply to every digitised curve.
Whitepaper
28 Mar 2023
18 pages · 17 min
A Reference Pipeline for Mask-to-Vector Curve Reconstruction
Most published well-log segmentation work stops at the mask. It reports an IoU or an F1 and treats the per-pixel probability map as the deliverable. For an operator it is not the deliverable. A petrophysicist cannot load a pixel mask into an interpretation package, cannot cross-check it against an LAS file, and cannot run petrophysics on it. The thing they need is a curve: one value per depth, monotone in depth, exportable as a column of numbers. The gap between the mask and that curve is small, unglamorous, and decisive, and it is where a digitisation system either earns the operator's trust or loses it. This whitepaper is the reference pipeline for that gap in VeerNet, the raster well-log digitisation architecture we built at EarthScan. We take the network's three-class mask, background plus two curve classes, and walk it through four deterministic stages: extract the per-depth centreline of each curve region so a band of foreground pixels collapses to a single ordinate; convert that pixel ordinate into a calibrated petrophysical value against the track's printed scale; fit a cubic spline through the resulting depth-value pairs so gaps from faint or broken traces are bridged smoothly rather than left as holes; and resample the spline to a fixed grid of 300 depth points that exports cleanly as CSV. We report the reconstruction errors we actually measured against held-out ground-truth curves under the Dice-trained model: a CSV mean absolute error of 0.11 for the first curve and 0.12 for the second, with mean squared errors of 0.03 and 0.04 respectively, on the model's normalised value range. We cover why the centreline, not the mask boundary, is the correct estimator; why cubic-spline interpolation beats linear for a physically smooth log response; why 300 points is a deliberate resolution choice rather than an arbitrary one; and where the residual error comes from so an interpreting team knows exactly what they are reviewing. The segmentation is upstream of this document. What follows is the deterministic post-processing that turns a mask into a number a petrophysicist can use.
Whitepaper
21 Feb 2023
19 pages · 20 min
Synthetic Data as Infrastructure for Industrial Computer Vision
Most accounts of synthetic data stop at the dataset: you generate some images, you train on them, you report a number. For an industrial computer-vision programme that has to serve a large and heterogeneous real archive, that framing is too small, because it treats the generator as a fixed input rather than as the part of the stack whose reach determines everything downstream. This whitepaper makes the platform argument instead. We treat the procedural log generator behind our raster well-log digitiser as infrastructure: a standing capability whose coverage of three axes, image size, curve count and scan noise, is a reusable asset that a fixed model draws on to generalise across the whole archive. The archive it had to serve is 136,771 TIF scans and 7,781 LAS files; the corpus that carried it was 15,000 synthetic training curves in the version-two build, drawn from a 20,000-log two-curve generator run, with images ranging 3,200 to 12,800 pixels wide and 480 to 640 tall, evaluated across 5 loss functions, and it contained zero human annotations. The claim we defend is that this is not a dataset you rebuild per job. It is a capability whose coverage sets an upper bound on the reach of any fixed model trained on it, whose one-time build cost amortises across every real scan the model ever serves, and whose breadth, not the count of real labels, is what carried the digitiser to a peak reconstruction R-squared of 0.9891. We separate this from the two questions it is often confused with, the internal design of the generator and the architecture of the model, and argue the platform view directly: coverage is the asset, reuse is the economics, and coverage-driven generalisation is the mechanism. The deliverable is a way of budgeting, staffing and defending a synthetic-data capability as durable infrastructure that any team standing up industrial CV under a no-labels constraint can reuse.
Whitepaper
20 Feb 2023
17 pages · 18 min
Training on One Image at a Time: solving the variable-dimension batch problem in log segmentation
Almost every published segmentation result quietly assumes that training images can be stacked into a rectangular batch tensor. Scanned well logs break that assumption at the source: a raster archive contains images spanning from roughly 3,200 to 12,800 pixels wide and 480 to 640 pixels tall, in no fixed ratio, because the physical logs were printed at different scales, photographed at different times, and scanned by different operators across decades. You cannot stack tensors of different shapes, so the naive training loop fails before the first forward pass. This whitepaper is a pure engineering account of how we trained VeerNet, our encoder-decoder segmentation network with a transformer attention bottleneck, under exactly this constraint. The binary-segmentation stage was forced to a batch size of 1, which trains correctly but starves the optimiser of the gradient averaging that makes large-batch training stable, and it took 110 minutes for 50 epochs over 2,000 instances. For the harder three-class multiclass stage on 15,000 synthetic instances we engineered a custom PyTorch collate function that pads a mini-batch to the largest member, carries a validity mask so padding never contributes to the loss, and pairs it with gradient accumulation to reach an effective batch of 16. That run completed 50 epochs in 550 minutes. We cover the memory arithmetic that sets the real ceiling, why padding-and-masking beats resizing for one-pixel-wide curve traces, the GroupNorm choice that keeps normalisation honest at batch size 1, and the throughput numbers an engineer needs to plan a training budget. The geoscience is upstream of this document; what follows is the systems problem of getting a variable-dimension dataset through a fixed-shape training loop without corrupting the gradient.
Whitepaper
17 Jan 2023
19 pages · 20 min
Class Imbalance in Thin-Structure Segmentation: A Field Guide
Under severe foreground scarcity, choosing a segmentation loss is only one of at least three decisions a practitioner has to make, and treating it as the whole answer is the most common way thin-structure digitisers underperform. This whitepaper is a field guide to imbalance mitigation for a target that occupies under two percent of the pixels, drawn from building VeerNet, our encoder-decoder segmentation network for raster well-log digitisation, where a curve trace is frequently one pixel wide and roughly ninety-seven percent of every log image is background. We survey three families of remedy rather than ranking objectives against each other. The first is loss weighting: a weighted binary cross-entropy with a positive-class weight of forty-two, set by the imbalance ratio, which lifts the binary-stage curve recall to a sourced ninety-six and ninety-seven hundredths. The second is region-based losses, the Dice and Tversky family evaluated among five objectives, which optimise overlap directly and are the run where both curve overlap and curve precision are measured together: on the multiclass Dice stage the curve intersection-over-union sits at twenty-six and twenty-one hundredths against a background at ninety-four, and the curve precision on that same run is a low forty-one and thirty-six hundredths. The third is synthetic balancing: procedurally generating logs that raise the genuine foreground share the optimiser trains against, which pays back in precision by giving the model real positives to learn from rather than reweighted copies of the few it has. The central argument, made visible in an interactive frontier board, is that each family moves a different axis and none moves all three, so a usable curve is a stacking decision, not a selection decision. We close with the operating posture we adopted, the peak intersection-over-union of fifty-one hundredths the combined stack reached, the failure modes each remedy introduces, and the order in which a team facing the same scarcity should reach for them.
Whitepaper
15 Nov 2022
22 pages · 21 min
Sim-to-Real Transfer for Document-Style Geoscience Images
The standard way to build a document-image model is to scan documents, label them, and learn from the labels. Raster well-log digitisation cannot follow that recipe, because nobody ever labelled which pixels in a scanned log belong to which curve, and labelling enough of them by hand is the exact manual cost the project exists to remove. Our alternative was to train end to end on a corpus that contains no real document at all: 20,000 procedurally generated logs from which a final version-two training set of 15,000 curves was drawn, every pixel mask manufactured by construction rather than annotated, and the model never shown a single scanned page during training. The wager in that choice is sim-to-real transfer, the question of whether a network raised in a generator survives the distribution shift to real scans. We argue it does, and we argue why. Evaluated on reconstructed curves against the native Log ASCII Standard data of a public Texas archive of 136,771 scanned TIF documents and 7,781 LAS files, the same fully synthetic training run reaches a peak coefficient of determination of 0.9891, with the model encountering those real scans for the first time at inference. The contribution of this document is not the headline number but the design discipline behind it: which axes of the synthetic corpus had to bracket the real archive's range for transfer to hold, why breadth of randomisation rather than photorealism is what carries the jump, where the reality gap reopens on degraded scans, and how we measured all of it so the claim that transfer survived is auditable rather than asserted. We present three interactive readers built directly on the engagement figures, walk the corpus-design decisions axis by axis, and state the operating rule we now apply to any document-style transfer problem of this shape.
Whitepaper
19 Oct 2022
19 pages · 20 min
Serverless Inference for Heavy Vision Models on AWS Lambda
Serving a heavy vision model on demand runs straight into a packaging problem that most model-serving writing skips, because most model-serving writing assumes a long-lived container with the model already resident in memory. AWS Lambda gives none of that for free: the deployment package is bounded, the layer is bounded, the runtime ships a deliberately minimal set of native tools, and the function is torn down between invocations. A raster-log digitiser whose trained checkpoint runs to hundreds of megabytes, whose dependency set is the full scientific-Python stack, and whose input is a single-channel grayscale scan up to 12,800 pixels wide, breaches every one of those limits at once. This whitepaper documents the packaging architecture we shipped to serve that model behind Lambda, and it makes one claim its central spine: the model cannot be served as a single deployment package, so it is decomposed into three artefacts, each of which exists to clear a different packaging constraint. The checkpoint is mounted from an Elastic File System, which removes it from the deployment-size accounting entirely and lets a function hydrate weights far larger than any package allows. The Python dependencies are published as a layer uploaded through S3, which keeps the function image small while still carrying a heavy import graph. An ImageMagick binary is shipped as its own layer, because the preprocessing that turns a raw grayscale scan into the single-channel tensor the model expects needs a native tool the Lambda runtime does not include. We set out the three limits explicitly, walk the decomposition that answers them, and describe the cold-start budget a batch-of-one request pays to hydrate the checkpoint from EFS and import the dependency layer, a cost paid only when a request actually arrives. Against that on-demand cost we place the fixed cost the posture avoids: an always-on GPU at 750 to 1800 EUR per month that bills whether idle or busy. The deliverable is a reusable packaging pattern for any team that has to serve a model too heavy for a function package, on infrastructure that only bills for the work it does.
Whitepaper
20 Sept 2022
18 pages · 17 min
Designing a Procedural Log Generator: Geometry, Annotations and Domain Randomisation
Almost every published segmentation result starts from a labelled dataset. Raster well-log digitisation does not get to: the archive is decades of scanned paper at every scale and condition, and nobody ever sat down to annotate which pixels belong to which curve. There is no corpus to learn from, and hand-labelling one at the scale a segmentation network needs is the exact bottleneck the project exists to remove. Our answer was to manufacture the dataset with a procedural log generator, and this whitepaper is the design account of that engine treated as a system in its own right. We describe the geometry layer that synthesises curve traces directly, so that a pixel-exact mask is produced by construction rather than by annotation; the annotation layer that stamps the grid lines, depth ticks, track separators and printed labels a real scanned log carries, so the network learns to ignore them; and the domain-randomisation layer whose knobs - image width from 3,200 to 12,800 pixels, height from 480 to 640 pixels, the number of curves per log, a single grayscale channel, and a tunable noise and realism level - sweep every render across the same distribution the real archive spans. The design discipline is that breadth of randomisation, not volume of hand labels, is what makes the dataset cover the problem: a generator dialled wide manufactured a 20,000-log two-curve multiclass corpus, and the final version-two training set drew 15,000 curves from it, none of them touched by a human annotator. We cover why synthesising the mask beats annotating one for a one-pixel-wide target, how the annotation layer is chosen to be adversarial to the model rather than decorative, which randomisation knobs matter and why each one is bounded by a real archive statistic, and the discipline that keeps a synthetic corpus from teaching the network the generator's own habits instead of the geoscience.
Whitepaper
19 May 2022
19 pages · 20 min
Choosing a Segmentation Loss Under Severe Foreground Scarcity
Choosing a segmentation loss is usually framed as a benchmark question: run an ablation, rank the candidates, ship the top scorer. We argue that for a thin-structure digitiser the framing is wrong, because the candidates do not differ on a single scalar and the right choice depends on which failure mode the product can absorb. This whitepaper is a first-person decision framework for selecting a loss under severe foreground scarcity, drawn from building VeerNet, our encoder-decoder segmentation network for raster well-log digitisation. The setting is extreme imbalance: the curve trace a model must recover is frequently one pixel wide, so at the engagement operating point roughly 97 percent of pixels are background, a ratio the weighted binary loss had to answer with a positive-class weight of 42. We evaluated five losses, Dice, Focal, Lovasz, Soft cross-entropy, and Tversky, and rather than reporting a single leaderboard we lay out the three criteria that governed the call: recall priority, because a missed thin curve cannot be recovered downstream; calibration and the asymmetry of the penalty, because the cost of a false negative on a thin curve far exceeds the cost of a false positive; and the propagation of the loss choice into the deliverable, the mean absolute and mean squared error left on the digitised curve a petrophysicist actually opens. On the mask the candidates are nearly indistinguishable, with curve-class F1 clustered around 0.32 to 0.37; on the recovered curve they separate cleanly, with Tversky leaving the lowest curve-1 mean absolute error of 0.0277 against Dice at 0.0367 and Focal at 0.0405, and the peak curve-1 coefficient of determination of 0.9891. We embed three interactive instruments, a five-loss decision matrix, a foreground-scarcity class-weight dial, and a downstream-error scatter, and close with the operating rule we adopted: under this regime, pick the loss whose penalty asymmetry you can tune to the failure you fear, grade it on the curve it emits, and treat the class weight as the first design decision rather than a hyperparameter to sweep.

Tell us about your data — we'll come back within one working day

Three short steps. We use your answers to route to the right person on our side and prepare a relevant first reply.

Continuous AI for explorers

info@earthscan.io

About Earthscan

Products

Legal

Follow us on

Go to Top

Whitepapers

Why General LLMs Fail at the Wellbore: The Case for Domain-Native AI in Petroleum Geomechanics

Hulde at the Wellbore: Domain-Native AI That Earns Engineering Trust

Own Your AI Stack: A Whitepaper for Energy Operators

Running AI R&D Like a Board-Grade Program: RAID Logs, GDPR Rails, and Three-Audience Reporting

VeerNet: AI for Raster Well-Log Digitisation

MIMEA: AI for Biomass and Bio-Inspired Materials

Lead KPIs, Lag KPIs, and a Bi-Weekly Verification Cadence: Governing an AI R&D Programme for Handover

Anatomy of a Multi-Phase, 20-Month Applied-AI Engagement: How the Money and Milestones Were Structured

Papering an AI Research Engagement: Effort Obligations, Phase-Capped Liability, and a Majority IP Split

Anatomy of a Fixed-Price AI R&D Phase: 250 Mandays, Five Roles, and the Discount Math

A Full-Stack Carbonate Subsurface-AI Reference Architecture: Fractures, Beddings, Vugs and Well-to-Well in One Pipeline

Contracting for Uncertainty: How to Paper an AI R&D Engagement

After the Model Ships: Anatomy of a Year-Two AI Technology-Transfer Engagement

Building Sovereign Subsurface-AI Capability: From an R&D Engagement to In-House Continuous AI

The Hardware Moat: Why Owning GPUs Beat Renting Them Through a Macro Shock

After the R&D Ends: Three Ownership Models for the AI You Just Built

From Manual Picks to a Continuous Interpretation Workflow: Operationalizing AutoFrac and AutoVug for Reservoir Engineering

Compute Budgeting for Vision Training Under a Hard GPU-Memory Ceiling

Deploying Subsurface AI Inside a Closed NOC Network: The On-Prem MLOps and Geo-Platform Integration Blueprint

Anatomy of a Top-K Mode-Subtraction Vug Detector: Five Enhanced Copies of One Log

One Classical-CV Parameter Set, Every Well: A Field Guide to the Eight Knobs of the Vug Detector

Recovering Trapped Data: A Five-Year View of Industrial Digitisation

Vug Fingerprints, Not Vug Percentages: Per-Interval Morphology as a Reservoir-Quality Log

From 3-4 Weeks to 2 Hours per Well: The Economic Case for Automating Borehole-Image Interpretation

Data-Centric AI Under Extreme Well Scarcity: How 235 Patches Became 75,000 and Why 14 Wells Beat Synthetic Overload

From One-Off to On-Demand: Productising a Raster-Log Digitiser into a Hosted Scan-to-LAS Service

MLOps for Subsurface AI: Reproducibility, Versioning and Handover

Half the Bits, Twice the Headroom: Mixed-Precision Training for 12800-Pixel Log Images

Well-to-Well Domain Adaptation: Carrying a Fracture and Bedding Model Across Wells with Kriging and Stratigraphic Correlation

A Phased Operating Model for Standing Up Subsurface AI Capability

Beyond Bulk Percentages: AI-Assisted Individual Vug Quantification in Carbonate Reservoirs

Unit Economics of a Curve-Digitisation SaaS for Subsurface Teams

AI for Long-Term Energy Project Engagement and Stakeholder Management

GeoBFDT: End-to-End Detection Transformers for Fracture and Bedding Picking in Carbonate Image Logs

Evaluating Curve-Extraction Quality Beyond IoU and F1

An MLOps Blueprint for Vision Models in Resource-Constrained Industry

Missing-Data Diagnostics Before Petrophysical Machine Learning

Encoder-Decoder Backbones for Variable-Aspect-Ratio Log Images

Sub-Pixel Centrelines and Depth Registration: A Validation Framework for Digitised Curves

A Reference Pipeline for Mask-to-Vector Curve Reconstruction

Synthetic Data as Infrastructure for Industrial Computer Vision

Training on One Image at a Time: solving the variable-dimension batch problem in log segmentation

Class Imbalance in Thin-Structure Segmentation: A Field Guide

Sim-to-Real Transfer for Document-Style Geoscience Images

Serverless Inference for Heavy Vision Models on AWS Lambda

Designing a Procedural Log Generator: Geometry, Annotations and Domain Randomisation

Choosing a Segmentation Loss Under Severe Foreground Scarcity