AI-assisted graded flux laser and ATIG welding of P91 steel and AISI 904L dissimilar joints for supercritical power plant piping
Pressure vessels and boiler systems operating under extreme thermal and corrosive loads often need a transition joint between two very different materials: one chosen for creep strength at high temperature, the other for resistance to aggressive chloride or sulphate attack. P91, a modified 9Cr-1Mo ferritic-martensitic steel, is widely specified for creep performance in high-temperature piping. AISI 904L, a super-austenitic stainless steel with high Mo and Cr content, is chosen where sulphate- or chloride-rich environments would otherwise cause rapid degradation.
Joining these two is rarely attempted, for good reason: P91 is prone to hydrogen-assisted cold cracking, and 904L's high alloy content makes it prone to Mo/Cr segregation and secondary phase formation under uncontrolled thermal cycles. No prior literature reports a flux-assisted or AI-optimized approach for this specific combination. This disclosure proposes a multi-oxide activated flux, a TiO₂-SiO₂-Cr₂O₃ blend with a minor Fe₂O₃ addition, applied as a paste before autogenous laser welding, with the flux composition optimized computationally before any physical trial is run.
The optimization runs as a two-stage loop. First, a dataset assembled from published flux-assisted weld studies trains a regression model (Gradient Boosted Trees or Random Forest) to rank candidate TiO₂-SiO₂-Cr₂O₃-Fe₂O₃ ratios by predicted joint quality, narrowing a large compositional search space to a short, physically testable list. Second, once real samples exist, a convolutional neural network trained on metallographic cross-sections converts each image into a quantitative defect and quality score, feeding back as ground-truth labels that sharpen the regression model's next round of predictions. Bayesian optimization decides which untested composition to try next, prioritizing whichever candidate the model is currently most uncertain about, rather than testing exhaustively.
The flux itself is applied in a spatially graded arrangement: a Cr₂O₃-Fe₂O₃-enriched formulation toward the P91 side, a TiO₂-SiO₂-enriched formulation toward the 904L side, overlapping gradually near the fusion interface. That grading lets the flux chemistry match each parent metal's different metallurgical needs within a single weld pass.
Every predictive chart in this disclosure, grain size, tensile correlation, joint-quality ranking, is explicitly illustrative rather than experimentally validated: this patent describes a proposed methodology and the AI framework meant to guide the physical welding trials, not completed results. The physical welding, metallographic, mechanical, and hot-corrosion validation is the next phase of this work.
