Toward an AI-driven materials modeling platform

Important results from the project

This project enabled us to acquire the knowledge needed to implement modern machine learning (ML) models applicable to material property predictions. We showed the ability of ML models trained on company data to deliver useful predictions on new material systems. In addition, with the application of Bayesian inference, we were able to quantify the uncertainty in predictions made with our software. This functionality is required to fully integrate a thermodynamic modeling program with modern process and component design simulation tools.

Expected long term effects

The project implemented machine learning (ML) pipelines for selected thermodynamic properties of two classes of materials that are typically modeled with the Thermo-Calc package obtaining ML models with predictive ability. For the project, this is the beginning of the journey of applying ML to a range of properties and classes of materials of interest to our customers. Furthermore, the Bayesian uncertainty quantification software we developed will be the framework from which we will start to include in the package the uncertainty quantification functionalities that our customers are requesting.

Approach and implementation

The project implemented the ML pipelines in Python using industry-standard libraries (e.g., scikit-learn, TensorFlow, and PyMC3) and built a MongoDB database to manage the data. The project prototyped a Python module for Bayesian uncertainty quantification that uses TC-Python, our proprietary Python API to the Thermo-Calc package, and an efficient Markov-Chain Monte Carlo algorithm. To disseminate knowledge in the company, we held an internal workshop with lectures on ML by the external expert and presentations on the results of the project.