Combined CALPHAD and Machine Learning for Property Modelling

Abstract: Techniques to improve the speed at which materials are researched and developed has been conducted by investigating the machine learning methodology. These techniques offer solutions to connect the length scales of material prop- erties from atomistic and chemical features using materials databases generated from collected data. In this assessment, two material informatics methodologies are used to predict material properties in steels and nickel based superalloys using this approach. Martensite start temperature and sigma phase amount as a function of input composition has been modelled with the use of machine learning algorithms. The experimental methodology had a collection of over 2000 unique experimental martensite start temperature points. This yielded important information on higher order interactions for the martensite start temperature, and a root mean square error (rmse) of 29 Kelvin using ensemble tree based algorithms. The metamodel was designed using an artificial neural network from TensorFlow’s library to predict sigma phase fraction and its composition. The methodology for building, calculating, and using data from TC-Python will be laid out. This generates a model that would generalize sigma phase fraction 97.9 % of Thermo-Calc’s equilibrium model in 7.1 seconds compared to 227 hours neded in the simulation to calculate the same amount of material property data.