Discovering chemistry to creep rupture equations in Alloy 617 with machine learning

This study seeks to discover a mathematical relationship between stress, temperature, properties, chemistry, and creep-rupture of a superalloy. This discovery will be achieved by leveraging human-supervised machine learning (ML). Historically, creep rupture equations have been discovered based on human analysis of experimental data. Numerous equations have been derived that describe the relationship between stress, temperature, and creep-rupture; however, a mathematical relationship with chemistry has remained outside the domain of human understanding. Recent advancements in ML offer the opportunity to discover equations with a higher dimensionality. To that end, multigene genetic programming (MGGP) with symbolic regression; a biologically inspired ML method, is employed to derive human-interpretable creep-rupture equations for Alloy 617. The optimal equation is observed to be a function of stress, temperature, chemistry, and chemical ratio in a mathematical form that corresponds to creep-strengthening/weakening mechanisms. The predictions agree statistically with the creep data including blindly held data for post-audit validation. The equation is leveraged to discover an optimal chemistry for Alloy 617 that offers improvement in creep strength. Calculated equilibrium phase diagrams (CALPHAD) of the optimal chemistry show an increased phase fraction of strengthening carbides that are stable over a wider temperature range.