Uncertainty Driven Computational Thermodynamics

In computational thermodynamics, statistics and uncertainty quantification are critical for evaluating the confidence in predictions and models with substantial potential for impact on materials development and qualification. The role for statistical techniques, especially Bayesian analysis, is not limited to obtaining uncertainty intervals on phase boundaries. These methods can drive the development of selection criteria for the best combinations of datasets and thermodynamic descriptions that support each other for a given material system. We present the results of such uncertainty driven investigations into thermodynamic model development for unary and binary metallic systems. Specifically, we share statistical approaches to Bayesian model selection, parameter inference, uncertainty quantification/propagation, and the automated weighting of both experimental and computed datasets. This will be accompanied by a discussion of the use of statistical techniques in thermodynamics to inform critical design choices and gleaning scientific insights.