Technical evaluation · for E&P geophysics engineers
SEC-grade audit trail for the subsurface model.
For processing geophysicists and reservoir engineers at operators and service companies. The same pipeline that channel-aware receivers use in telecom applies to seismic: extract the subsurface state M t from the trace via Wiener spiking deconvolution, track it across 4D monitor surveys with a constant-velocity memory hop, and seal each M t into an HMAC transition chain bound to an FPGA-resident PUF key. A single bit altered in any historical survey breaks every downstream signature. Every figure was produced by the published Python pipeline — industry-standard convolutional model, no proprietary data.
Request engineering briefing Read the whitepaper
§0 · Reader's note — what is and is not claimed
Two independent value props on this page. (1) is deliberately standard practice.
(1) Wiener spiking deconvolution + memory-hop tracking (§3 – §4). This is what every E&P processing shop already does — Wiener deconvolution against a known source wavelet is textbook seismic processing (Yilmaz Ch. 2). The reconstruction quality in §3 is not a novel processing claim. The licensable contribution is the chain-anchored α-β filter — every 4D monitor survey's M t is tracked against a chain-trusted base survey, so retroactive edits to the historical baseline are mechanically impossible. Pilot validation on real proprietary 4D monitor surveys is the next step.
(2) SEC-grade audit trail (§5 – §7). The FPGA-PUF sealed transition hash chain over multi- decade seismic acquisition is the standalone licensable asset. SEC reserves attestation, SOX §404 controls evidence, consortium data-sharing under PSC / JOA terms, and contractor-vs-operator liability disputes all need a tamper-evident provenance of the seismic record. The SEG-Y user-trailer embed (XC5) means existing acquisition / processing / interpretation tools pass it through unmodified. Applies regardless of which deconvolution / migration / AVO pipeline produced M t .
A processing geophysicist sceptical of (1) on "I-already-do-this" grounds can still adopt (2) standalone — the chain primitive is processing-agnostic.
⚡ Signal / Data Restoration
Noisy seismic survey → subsurface medium state reconstructed.
Patent core claim: "recover original modal information from broken received signal via medium forward model + integrity-anchored prior" . Wiener deconvolution with wavelet + Q-attenuation kernel as M t , combined with α-β filter tracking against a chain-verified base survey, reconstructs reservoir state from noisy 4D monitor surveys — worst-step error 15.5 → 5.2 ms. SEG-Y user-trailer embed, no upstream changes.
§1 · Why an upstream operator should read this
Two value axes: better recovery, and audit-grade history.
Wiener spiking deconvolution against a known source wavelet is well-established processing — the recovery gain in §3 is what every E&P shop already does. The new piece is the chain: every M t (the reservoir-state descriptor — reflector amplitudes and two-way times) is signed by a PUF key inside an FPGA and bound to its predecessor. The result is a cryptographically-anchored seismic record that survives contractor handoff, consortium sharing, and long-term archive — and that can be independently verified by regulators and partners.
SEC reserves attestation
Tamper-evident provenance of the seismic data underlying booked reserves. Compatible with SOX §404 controls.
Consortium data sharing
Non-operating partners verify the same M t chain offline with a public verification key — neutral evidence base.
4D reservoir monitoring
Memory-hop tracking of the reservoir front through multiple monitor surveys, with each survey integrity-sealed.
Contractor-vs-operator liability
Sealed snapshot at every processing step — acquisition contractor handoff, processor, interpreter — supports dispute resolution.
§2 · Experimental setup
1-D convolutional model — the geophysics-textbook baseline.
Standard convolutional model t(τ) = w(τ) ⊛ r(τ) + n(τ). A 6-reflector layered earth includes a moving reservoir-top reflector that migrates 60 ms across 8 monitor surveys (typical pressure-front response in a waterflood). A 30 Hz Ricker source is the canonical land-seismic wavelet; sampling and trace length are field-typical.
Earth model
6 reflectors, two-way times 0.30 / 0.55 / 0.80 / 1.10 / 1.45 / 1.60 s; reservoir top at 1.45 s migrates to 1.39 s over 8 surveys
Source wavelet
30 Hz Ricker, zero-phase, 251-sample length (typical land seismic)
Sampling
2 ms (Δt), 1001-sample trace (2 s record)
Noise
White Gaussian on the receiver trace; SNR specified in dB relative to the noiseless trace power
Sample size
32 independent noise realizations per SNR point (single-survey sweep); 24 per 4D step
Receiver A — no M t
Raw trace, no deconvolution. Reservoir-top picked from the largest |amplitude| event in the reservoir zone (1.20–1.70 s).
Receiver B — with M t
Frequency-domain Wiener spiking deconvolution using the known wavelet (the wavelet + Q-attenuation kernel is M t for this medium)
Receiver C — memory hop
Wiener deconvolution + α-β filter on the verified-past reservoir position and velocity; anchored search around the predicted location
§3 · Single-survey reflectivity recovery
Wiener spiking deconvolution — the standard processing gain.
Receiver B is textbook Wiener deconvolution. We reproduce its expected gain over the raw trace as a baseline that any processing geophysicist can verify against their own tools (Petrel, Kingdom, OpendTect, ProMAX, Madagascar). The point of this section is not novelty — it is to establish the floor that §4 and §5 build on.
Input trace SNR
Receiver A — ρ (raw)
Receiver B — ρ (with M t )
−5 dB
0.215
0.308
+0.093
0 dB
0.316
0.378
+0.062
5 dB
0.392
0.422
+0.030
10 dB
0.426
0.456
+0.030
15 dB
0.442
0.487
+0.045
20 dB
0.446
0.513
+0.067
ρ = Pearson correlation between recovered reflectivity and ground-truth spike series (industry-standard deconvolution score). 32 independent noise realizations per SNR point. The gain is modest at high SNR (raw trace already wavelet-aligned to the reflectivity locations) and largest at low SNR — exactly the regime where field acquisition typically lives.
§3.5 · 4-class saturation classification under field noise
Field crews care which kind of change happened, not just whether it did.
Beyond reflectivity recovery (§3), the operator needs to classify the change in reservoir state — no change / minor / moderate / major (production sweep) — so that asset-management and SEC reserve reporting can act on it. Receiver A is the cross-correlation method most current interpretation toolchains use. Receiver B projects the reservoir window onto a modal basis + cross-correlation lags + spectral bands (16-dim feature vector). Both use LDA on the same training set; only the feature representation differs.
Fig. — Modal + sealed prior delivers +20 to +24 percentage points over the cross-correlation baseline across every SNR. Even at −10 dB ambient + multiples + surface-wave, B still reads at 70.4% vs A's 47.3%.
SNR (dB)
Receiver A — 3-band cross-correlation
Receiver B — modal + sealed prior (16-dim)
73.9%
98.3%
+24.4 pp
74.6%
95.5%
+20.9
67.8%
89.2%
+21.5
56.8%
76.7%
+19.9
47.3%
70.4%
+23.1
Fig. — Clean seismic traces, 4 saturation-change classes. The reservoir window (~1.0 – 1.2 s) is where the class-distinguishing reflectivity changes occur.
Fig. — Moderate-change trace at SNR = −5 dB: ambient + multiples + 8/14 Hz surface waves dominate the view. Receiver A's cross-correlation collapses; Receiver B's modal features survive because each band integrates over class-specific spatial-frequency content.
Numbers from running the published Python pipeline: simulated 4-class seismic traces (35 Hz Ricker wavelet, 3-layer reservoir, class-dependent saturation perturbations), corrupted by ambient + multiples + surface-wave. 400 traces per class per SNR. Both classifiers use LDA on an SNR-mixed training set. Reproducible from fixed seeds.
§4 · 4D monitor surveys · memory hop
Verified past constrains the present. The reservoir front gets tracked, not chased.
The reservoir top migrates 60 ms in two-way time over 8 monitor surveys — a representative waterflood pressure front. Acquisition noise (SNR = −5 dB per survey) makes every individual survey unreliable. Without history, a picker globally searches the reservoir zone and gets fooled by noise peaks. With a chain-verified prior, an α-β filter predicts the next position from history, and an anchored search rejects noise outside the prediction window.
predict : τ̂ k|k−1 = τ̂ k−1 + v̂ k−1 residual : r k = z k − τ̂ k|k−1 correct : τ̂ k = τ̂ k|k−1 + α·r k , v̂ k = v̂ k−1 + β·r k
Receiver
Avg reservoir-top error
Worst step
Behavior
A — raw trace (no M t )
10.22 ms
22.0 ms
Noise peaks often beat the reservoir
B — memoryless M t
5.82 ms
15.5 ms
Deconvolution helps but no temporal context
C — memory hop M t
2.38 ms
5.2 ms
α-β predicts; verified prior rejects outliers
Memory hop achieves 2.44 × lower average error than the memoryless Wiener — and the worst-step error is roughly a third. Worst-step accuracy is what matters when a single misidentified reservoir position can derail a production-allocation decision.
§5 · Transition hash chain · SEC-grade audit
Each M t is signed by an FPGA-resident PUF key, then chained.
Memory hop in §4 is only as trustworthy as the past it leans on. If a contractor or partner could rewrite an earlier survey's M t , the receiver's prior is poisoned and the booked reserves derived from it are wrong. Subsystem C hashes each M t into a transition chain — every state HMAC-signed against the previous transition hash, with the MAC key derived from a PUF inside an FPGA. A 5 % amplitude tamper to one survey breaks every downstream link.
Fig. — 8-survey chain. A 5 % amplitude change to M 4 in storage: 4 / 8 links validate (clean baseline = 8 / 8). All H k for k ≥ 4 fail.
σ t = HMAC K PUF (H t−1 ‖ M t )
H t = SHA256(H t−1 ‖ M t ‖ σ t )
verify HMAC K PUF (H t−1 ‖ M t ) ≟ σ t ∀ t
MAC primitive
HMAC-SHA256, 256-bit PUF-derived key
Key custody
AWS F2 FPGA (AFI agfi-085919c09c35982c3 ) in the lab today; production target = a sealed rack-mount appliance at each processing site
Where it goes in the workflow
Inserted at every handoff: acquisition contractor → operator → processor → interpreter → reserves report. Each step adds its sealed M t to the chain.
Partner verification
Non-operating partners hold a public verification key; they can audit the chain end-to-end without ever seeing the PUF root
Tamper sensitivity (measured)
5 % amplitude change to one M k → 100 % of links from k to T reject
§6 · Where this pays back
Audit, settlement, and dispute resolution — at upstream-deal scale.
SOX / SEC §229
Sealed seismic provenance supports SOX §404 internal-controls assertion over proven reserves disclosed under SEC Reg S-K §229.1200.
Asset transactions
Buyer of a producing field can independently verify the seller's seismic chain before closing — replaces months of third-party reprocessing.
JV / consortium settlement
Operator publishes a chain; non-operating partners verify offline. Removes the "did the operator selectively share" class of dispute.
Contractor handoff
Acquisition crew countersigns the raw shot gather chain at vessel/truck handoff. Operator countersigns post-load. Liability boundary becomes cryptographic.
20-year archive
A 4D reservoir under monitoring has a 20-year history. Re-processing decades later with different algorithms still verifies against the original chain.
Sidecar appliance
Drops next to existing Petrel / Kingdom / ProMAX rigs — no change to the processing stack. M t is small (≪1 kB per survey) and the chain is content-neutral.
§7 · Integration paths
Three deployment shapes — none of them touches your processing software.
Sidecar appliance
A 1U / 2U server with an FPGA card sits next to the processing rig and seals each intermediate M t as it is written to disk. Petrel / ProMAX / Kingdom unchanged.
SEG-Y trailer chain
Sealed M t stored in a SEG-Y file's user trailer (or a side-car JSON). Any reader can verify; tools that don't care simply ignore it.
Cloud / Cyberinfrastructure
For operators running on AWS / Azure / OSDU — the FPGA seal runs in an enclave (Nitro / AMD-SEV) with the PUF emulated by a hardware secure module. Same chain primitive, cloud form factor.
Licensing model
Per-appliance hardware license, or per-TB seismic-volume seal-verification service. The lead financial advisor coordinates deal structure with the operator's commercial team.
§8 · Honest limitations
What this page does not show.
1-D convolutional model only. No NMO, no migration, no AVO, no anisotropy, no multiples, no surface waves. A production processing stack handles all of these; the chain primitive attaches downstream of whichever processing the operator runs.
Wiener deconvolution is not a new algorithm. Receiver B is textbook 1970s-era processing. The contribution of this page is the chain (§5), not the deconvolution. We reproduce §3 only to establish the floor every other claim builds on.
The 4D scenario is synthetic. A 60 ms reservoir-front migration over 8 surveys is a representative but stylized waterflood response. Real 4D monitoring sees AVO changes, time-lapse shifts, and interpretation noise we do not model here. Repeating §4 against the operator's own 4D dataset is the first useful pilot step.
FPGA is software-emulated for the chain in this demo. The 8-survey chain shown in §5 was computed in Python against the same PUF-key constant the FPGA uses. A separate program of work runs the same HMAC on a real AWS F2 FPGA AFI; wiring the processing pipeline to that FPGA is the production integration step.
Chain content vs chain attestation. The chain attests which M t was produced at which step. It does not encrypt the seismic data nor attest to interpreter identity. Confidentiality and identity binding are separate primitives.
α-β filter is illustrative. The §4 memory hop uses a simple constant-velocity α-β filter. A Kalman filter or particle filter that models pressure-front dynamics would tighten the tracking further, but is out of scope for this page.
§9 · For your geophysics team
Run it on your own data.
The Python pipeline that produced every figure on this page is available under NDA for evaluation. The most useful first step is to repeat §3 / §4 against your operator-specific 4D dataset (or against Marmousi / SEAM / Norne if you prefer a public benchmark) and report the same tables against your in-house processing chain. The chain primitive (§5) is data-independent and can be evaluated separately from the receiver work.
Request engineering briefing Read the whitepaper