Odin / SkyFrac
Reconstruction & Attack Plan
A technical consolidation of public claims and physics first-principles to define a legally clean independent equivalent, prove it works, and interrogate vendors.
The Core Thesis
Odin/SkyFrac is not computing exact "true fracture compliance" from surface data alone. Instead, they provide a real-time frac intelligence platform that infers hidden downhole hydraulic state.
What They Likely Do:
- Train models on rich historical, downhole, and full-sensor datasets.
- Deploy surface/minimum-sensor models during live treatments.
- Convert inferred state into a compliance-inspired "C-Factor" (frac-quality/restriction-risk metric).
What We Must Avoid Saying:
- • Do not claim they compute "exact true geomechanical compliance."
- • Do not state they definitely use PINNs, LSTMs, or XGBoost.
- • Do not claim they can uniquely invert true BHTP from surface data alone without uncertainty.
Inference Architecture
The Mathematics of Restriction
1. Fracture-Face Pressure
We must strip out wellbore, perforation, and tortuosity friction to find the pressure actually seen by the fracture.
Where wellbore pressure accounts for hydrostatic head and pipe friction:
2. Apparent Elastance (Restriction)
Instead of raw instantaneous derivatives, we use causal windowed arrived volume (\(\Delta V_w\)).
3. Apparent Restriction Index (ARI)
The final normalized operational metric.
The Flaw of Raw Compliance & Gates
The naive industry formula for compliance is \( C = Q / (dP/dt) \). It is operationally dangerous because pipe friction, rate changes, and phase lag pollute the derivative. Furthermore, any real engine must explicitly gate against these physical confounders:
The Missing Link: How Do They Predict 20 Mins Early?
Odin claims to see screenouts up to 20 minutes before surface pressure spikes. A purely physics-based \(dP/dV\) equation cannot do this. To beat them, we must build the Predictive ML Layer.
Hypothesis 1: Time-Series Anomaly Detection
Using sequence models (like LSTMs or Temporal Convolutional Networks) trained on the first derivative of the Apparent Restriction Index (ARI slope). The model learns the subtle trajectory curve that inevitably leads to failure before the human eye can spot the inflection point.
Hypothesis 2: High-Frequency Signatures
Looking at the acoustic/pressure reflection signatures (like subtle water hammer degradation) within the 1Hz or 50Hz data. Changes in the reflection time can indicate near-wellbore choking long before gross pressure rises.
The Validation Strategy
How we prove to leadership that our engine actually works, without needing to see Odin's code.
Test A: BHTP Reconstruction
Can our physics + residual ML reconstruct true downhole pressure on held-out stages? We must track RMSE and bias specifically during sand ramps and chemical changes.
Test B: ARI vs Raw Surface Pressure
Does our ARI detect restriction earlier and more reliably than a raw STP derivative? We will measure lead time, precision, recall, and false positives per stage.
Test C: Confounder Stress Test
Ensure the model does not confuse an FR loss with rock stiffening, or proppant duning with fracture-face restriction.
Test D: Production Correlation
Does our stage-level ARI score actually add explanatory power to EUR models compared to raw cluster counts and fluid intensity?
The Interrogation Playbook
If we are in a bake-off or meeting with vendors claiming "true compliance," ask these exact questions to reveal their true architecture:
- 1. "For a new basin, do you need a science well with downhole gauges before your surface-only deployment becomes reliable?"
- 2. "If Friction Reducer drops mid-stage, how does your model avoid misclassifying the pipe friction spike as reservoir stiffening?"
- 3. "Is your C-Factor computed from STP, estimated BHTP, or fracture-face pressure? How do you prevent it from blowing up when dP/dt is near zero?"
- 4. "Does '50 ms response' mean your actual ML model inference time, or just sensor-to-dashboard refresh latency?"
The Execution Plan
We are not trying to copy Odin's secret code. We are building a superior Apparent Restriction Index (ARI) Engine from first principles that is fully explainable.
Apply hydrostatic, residence-time, and friction corrections before attempting ML inference.
Explicitly classify and ignore transient states (water hammer, step-downs, FR drops).
Provide probabilistic restriction trajectories with reason codes, not deterministic black-box commands.