| Abstract Scope |
Characterizing creep for high-temperature components like turbine blades, engine nozzles, and reactor pressure vessels requires hundreds of tests extending up to years, bottlenecking material qualification. Predictive models are essential, yet empirical approaches extrapolate unreliably and mechanistic models remain tied to specific deformation mechanisms. Machine learning can build predictive models from data, but most approaches function as black boxes. Symbolic regression, an interpretable ML algorithm, addresses this by enabling discovery of mathematical expressions directly from data. In this study, a multi-population evolutionary algorithm is employed to derive a creep deformation equation from 304 stainless steel data spanning three temperatures and multiple stress levels. Candidates are selected from a Pareto front of accuracy versus complexity through human inspection for physical plausibility and analytical differentiability. The discovered equation achieves strong predictive accuracy on training data and shows excellent generalizability when applied to unseen data from P91 steel and IMI-834 titanium alloy. |