CorrosionAI
Request Demo

Physics-Informed Graph Neural Network for Corrosion Prediction

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.

Architecture Overview

Data Input Layer

Sensor telemetry, environmental conditions, material composition, and historical operational records from SCADA/historian systems.

Graph Neural Network Layer

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.

Physics-Informed Constraints

Butler-Volmer kinetics and Pourbaix thermodynamic diagrams embedded directly into the loss function, ensuring predictions are physically plausible.

Prediction & Uncertainty

Corrosion rate (mm/yr), remaining useful life (1-20 years), risk classification, and Bayesian Monte Carlo confidence intervals at 50%, 90%, and 95% levels.

How PI-GNN Works

1

Data Ingestion & Preprocessing

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.

2

Physics-Informed Constraint Layer

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.

3

Graph Neural Network Processing

Message passing across the graph allows each node to aggregate neighbor information. Attention mechanisms learn which spatial relationships are most informative for corrosion prediction.

4

Prediction & Uncertainty Quantification

Bayesian Monte Carlo dropout with 100 stochastic forward passes generates point estimates and calibrated prediction intervals. Epistemic and aleatoric uncertainty are separately quantified.

5

Auto-Calibration Feedback Loop

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.

Performance Benchmarks

MetricPI-GNN (CorrosionAI)Pure Data-Driven GNNRandom ForestLinear Empirical
R² (Coefficient of Determination)0.9760.9210.8470.623
MAE (Mean Absolute Error)0.043 mm/yr0.089 mm/yr0.127 mm/yr0.234 mm/yr
RMSE0.067 mm/yr0.134 mm/yr0.189 mm/yr0.312 mm/yr
MAPE (%)4.2%8.7%12.3%21.5%
Inference Time<50ms<45ms~120ms<10ms

Comparison with Industry Models

CapabilityCorrosionAI PI-GNNde Waard-MilliamsNORSOK M-506FreeCorp
Machine Learning Based
Physics-Constrained
Uncertainty Quantification
Real-Time Inference
Graph Topology Modeling
Continuous Self-Learning

Key Technical Differentiators

Physics Constraints

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.

Graph Structure for Spatial Correlations

Captures galvanic coupling, flow-accelerated corrosion at bends, cathodic protection shielding, and MIC colony propagation — phenomena that point-by-point models fundamentally cannot detect.

Bayesian Uncertainty Quantification

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.

Real-Time Inference in <60 Seconds

Optimized sparse matrix operations, INT8 quantization for edge inference, and streaming architecture deliver 240-480x speedup over traditional laboratory analysis.

Integration & Deployment

RESTful API

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"
  ]
}

SCADA & Historian Connectors

  • OPC UA
  • Modbus TCP/RTU
  • MQTT
  • Siemens S7
  • Allen-Bradley
  • Wonderware
  • Schneider Electric
  • Honeywell

Deployment Options

On-Premises

Full stack on customer infrastructure. Air-gapped option available for offshore platforms and environments with strict data sovereignty requirements.

Private Cloud

Dedicated cloud instance (AWS, Azure, GCP) with customer-controlled encryption keys and data isolation.

SaaS (Multi-Tenant)

Fully managed service with shared infrastructure and logical tenant isolation. Ideal for rapid deployment and pilot projects.

Frequently Asked Questions

What machine learning model does CorrosionAI use?

How does PI-GNN differ from traditional corrosion models?

How does CorrosionAI quantify prediction uncertainty?

What is the minimum H2S detection threshold?

Can CorrosionAI predict localized corrosion and pitting?

How does CorrosionAI integrate with SCADA systems?

See PI-GNN in Action

Request a demo to see how Physics-Informed Graph Neural Networks predict corrosion with >95% accuracy in under 60 seconds.

Request a Demo