Duct corrosion costs the global economy up to $2.5 trillion annually. Traditional approaches to corrosion prediction rely on simplistic empirical models that fail to capture the complex electrochemical dynamics at play in real industrial environments.
CorrosionAI takes a fundamentally different approach. Instead of treating corrosion as a black-box regression problem, we built a Physics-Informed Graph Neural Network (PI-GNN) that encodes five core electrochemical equations directly into its architecture:
Marcus Theory governs quantum electron transfer rates at metal-electrolyte interfaces. By embedding this into our activation functions, the network respects the fundamental physics of oxidation-reduction reactions.
The Arrhenius Equation captures temperature-dependent kinetics. Our model dynamically adjusts corrosion rate predictions based on operating temperature, something purely data-driven models struggle with when extrapolating beyond training ranges.
The de Waard-Milliams model accounts for CO2 partial pressure effects, critical in oil & gas environments where sweet corrosion dominates.
Pourbaix thermodynamics provides pH correction factors, ensuring predictions remain valid across the full range of aqueous environments.
Reynolds correlations capture flow-induced corrosion, where high-velocity fluids accelerate metal loss through erosion-corrosion mechanisms.
The result: validated against NIST and NACE experimental datasets, CorrosionAI achieves R² = 96.3% while maintaining full explainability. Every prediction can be decomposed into its contributing physical factors.
Our target sectors include CCS/CCUS, Oil & Gas, Cement, Maritime, and Offshore, with particular strength in detecting trace H2S concentrations at ppm levels — a scenario where traditional models are essentially blind.
The best AI doesn't replace human expertise — it amplifies it. CorrosionAI gives asset integrity engineers a tool that thinks in the language of electrochemistry, not just statistics.