Study on the composition design and corrosion resistance of zirconium alloys based on machine learning

The design of corrosion-resistant zirconium alloys for nuclear fuel cladding presents significant challenges, including reliance on empirical knowledge, lengthy experimental cycles, and difficulties in establishing quantitative composition-performance relationships. To address these challenges, this study developed a closed-loop machine learning framework ("data-model-explanation-design-validation") integrating 71 zirconium alloy corrosion datasets with 71 elemental physicochemical characteristics to construct an "alloy factor" feature space. A systematic comparison of ten algorithms identified the Long Short-Term Memory (LSTM) neural network as the optimal predictive model. Multi-step feature selection revealed that the combination of Miedema's valence electron cloud density, valence electron count, and melting enthalpy minimized the model error. The SHapley Additive exPlanations (SHAP) model analysis further correlates the predictions with the elemental physicochemical features. Guided by the optimized model, a four-step systematic screening process is carried out on ∼12,000 candidate alloy compositions under the constraint of total alloying element content below 5.0 wt.%, and two novel quaternary zirconium alloys, Zr-0.2Sn-0.8Nb-1.5Fe and Zr-1Sb-1.5Fe-0.1Cu, are developed and experimentally validated. Triplicate parallel autoclave tests show that the two alloys achieve average corrosion weight gains of 22.046 mg/dm² and 25.076 mg/dm² respectively after 3 days of exposure in 0.04 mol/L LiOH aqueous solution at 360 °C and 18.6 MPa, which are 45.02% and 37.47% lower than that of the Zircaloy-4 (Zr-4) benchmark alloy. These alloys exhibited superior short-term (3-day) and mid-term (14-day) corrosion resistance under the tested conditions, and SEM characterization confirms that their superior corrosion resistance originates from the dense and defect-free protective oxide film, demonstrating superior corrosion resistance.