CorrosionAI uses a PI-GNN that combines Butler-Volmer kinetics and Pourbaix thermodynamic diagrams with graph-based deep learning to predict corrosion rates with >95% accuracy in under 60 seconds.
Sensor telemetry, environmental conditions, material composition, and historical operational records from SCADA/historian systems.
4-layer graph attention network with 128-dimensional hidden states and 8 attention heads. Nodes represent sensor locations; edges encode spatial relationships and flow direction.
Butler-Volmer kinetics and Pourbaix thermodynamic diagrams embedded directly into the loss function, ensuring predictions are physically plausible.
Corrosion rate (mm/yr), remaining useful life (1-20 years), risk classification, and Bayesian Monte Carlo confidence intervals at 50%, 90%, and 95% levels.
Self-calibrating sensors sample on a 60-second cycle. Raw telemetry undergoes signal validation, outlier detection, feature engineering (saturation indices, CO2/H2S fugacity), and graph construction.
Computes Butler-Volmer electrochemical currents, Pourbaix regime classification, Nernst equilibrium potentials, and Faraday's law mass-loss conversion. Physics residuals are added to the training loss function.
Message passing across the graph allows each node to aggregate neighbor information. Attention mechanisms learn which spatial relationships are most informative for corrosion prediction.
Bayesian Monte Carlo dropout with 100 stochastic forward passes generates point estimates and calibrated prediction intervals. Epistemic and aleatoric uncertainty are separately quantified.
Every 60 seconds, predictions are compared against new sensor readings. Online Bayesian updates improve accuracy over time. Sensor drift and concept drift are automatically detected.
| Metric | PI-GNN (CorrosionAI) | Pure Data-Driven GNN | Random Forest | Linear Empirical |
|---|---|---|---|---|
| R² (Coefficient of Determination) | 0.976 | 0.921 | 0.847 | 0.623 |
| MAE (Mean Absolute Error) | 0.043 mm/yr | 0.089 mm/yr | 0.127 mm/yr | 0.234 mm/yr |
| RMSE | 0.067 mm/yr | 0.134 mm/yr | 0.189 mm/yr | 0.312 mm/yr |
| MAPE (%) | 4.2% | 8.7% | 12.3% | 21.5% |
| Inference Time | <50ms | <45ms | ~120ms | <10ms |
| Capability | CorrosionAI PI-GNN | de Waard-Milliams | NORSOK M-506 | FreeCorp |
|---|---|---|---|---|
| Machine Learning Based | — | — | — | |
| Physics-Constrained | ||||
| Uncertainty Quantification | — | — | — | |
| Real-Time Inference | — | |||
| Graph Topology Modeling | — | — | — | |
| Continuous Self-Learning | — | — | — |
Butler-Volmer kinetics and Pourbaix diagrams embedded in the loss function reduce training data requirements by 10x and guarantee physically plausible predictions, even under conditions not in historical data.
Captures galvanic coupling, flow-accelerated corrosion at bends, cathodic protection shielding, and MIC colony propagation — phenomena that point-by-point models fundamentally cannot detect.
100 stochastic forward passes quantify epistemic and aleatoric uncertainty. Prediction intervals achieve 96.1% coverage at the 95% confidence level, feeding directly into API 580/581 risk-based inspection planning.
Optimized sparse matrix operations, INT8 quantization for edge inference, and streaming architecture deliver 240-480x speedup over traditional laboratory analysis.
POST /api/v1/predict
Content-Type: application/json
Authorization: Bearer YOUR_API_KEY
{
"pipeline_id": "pipeline_001",
"timestamp": "2026-02-08T10:30:00Z",
"measurements": {
"temperature": 65.5,
"pressure": 1250,
"co2_partial_pressure": 0.45,
"h2s_concentration": 12.3,
"ph": 6.2,
"flow_velocity": 2.8,
"water_cut": 0.35
}
}
Response:
{
"corrosion_rate": 0.087,
"uncertainty": {
"lower_bound": 0.071,
"upper_bound": 0.103,
"confidence": 0.95
},
"risk_level": "medium",
"recommendations": [
"Monitor pH levels closely",
"Consider inhibitor injection"
]
}Full stack on customer infrastructure. Air-gapped option available for offshore platforms and environments with strict data sovereignty requirements.
Dedicated cloud instance (AWS, Azure, GCP) with customer-controlled encryption keys and data isolation.
Fully managed service with shared infrastructure and logical tenant isolation. Ideal for rapid deployment and pilot projects.
Request a demo to see how Physics-Informed Graph Neural Networks predict corrosion with >95% accuracy in under 60 seconds.
Request a Demo