Predicting Ti alloys properties : Machine learning as a bridge between experimental results and ab initio predictions
(2019) THE 14TH WORLD CONFERENCE ON TITANIUM — Location: Nantes (10.June.2019)
Files
MM-predicting-Ti2019-final.pdf
Restricted Access - Adobe PDF
- 3.19 MB
Details
- Authors
- Abstract
- Traditionally, new advanced engineering materials with specific properties requires processing, testing and characterization steps on a large search space, which is a tedious and expensive process. To overcome this issue, guidelines for material design have been introduced by many authors during the last decades, often based on computer calculation. As an example, Morinaga introduced the “d-electron design method” during the 1980s’ for the prediction of the phase stability in titanium alloys. Recently, this design method has been extended for the prediction of alloys exhibiting high work-hardening capabilities. As a basic principle, ab initio calculations can help in reducing the number of potential candidates. However, two types of difficulties arise in practice. Firstly, numerous calculations requiring huge computer resources are necessary when including complex (i.e. amorphous, heavily doped, …) or multi component random alloys. Secondly, the macroscopic properties of interest may be difficult, if not impossible, to apprehend through simulations in a direct way. We developed in this study a method to solve these difficulties based on the use of machine learning to connect bulk properties to simple calculations via a database of experimental results. The investigated properties can be any bulk property, going from the electrical or thermal conductivities to the type of deformation mechanisms occurring under loading. These properties are coupled to descriptors derived as linear combinations of microscopic properties calculated for single stoichiometry two-component model systems. This method was applied to develop predictor models for the type of mechanical deformation occurring in TRIP/TWIP titanium alloys. A database of 8 descriptors including 20 common doping elements was built. The space of microscopic descriptors was investigated to derive a relevant subset when removing strongly correlated ones. Based on this subset, various machine learning models were trained and investigated in order to validate the most appropriate one when comparing the predictions for newly obtained experimental results.
- Affiliations
Citations
Marteleur, M., Choisez, L., Jacques, P., Rignanese, G.-M., & et al. (2019). Predicting Ti alloys properties : Machine learning as a bridge between experimental results and ab initio predictions. THE 14TH WORLD CONFERENCE ON TITANIUM, Nantes.