Machine learning phase data for exploratory Calphad Assessments
As of today, most Calphad assessments are based on pre-existing thermodynamic descrip-tions. However, the chemical composition space is still relatively unexplored, mostly fo-cussed on elemental combinations relevant for metallic alloy production, Figure 1. Certainly, most of the unknown combinations will have to be filled at some point to enable modelling of functional materials as well as their recycling.
In this work, we present an easy solution to a major challenge in an assessment ‘from scratch’: the necessity to guess stability and presence of phases. An exploratory approach in combination with machine learning facilitates swift development of Calphad descriptions for experimentally underdetermined systems.
Using the aiMP database  based on the high throughput database of ab initio calculations of Materials Project  with more than 100’000 structures, we are able to supply initial guesses for over 3000 binary systems with an estimate of stable intermediate phases.
The thermodynamic models of these compounds are optimized through machine learning methods on the 4000+ compounds in the self-consistent SGTE and FactSage Pure Substance databases, yielding an S298 estimate with a mean error of 2.3 kJ/molK, and a formation enthalpy mean error of 12 kJ/mol. We will present the actual utilization of assessing an experimentally vague system using this purely computational data.
An online resource for all binary databases will be available free of charge through gtt-technologies.de/data/ and is fully compatible to 2nd generation Calphad databases. The databases can be used directly in the free programs FactSageEdu or ChemAppLight.
 A. Jain, S.P. Ong et al., APL Materials, 1(1), 2013, 011002.