GEMPROMISE
uclouvain |
Generative machine learning for combined process control and materials design
GEMPROMISE aims to tackle the grand challenge of materials science, namely to identify the process parameters leading to a structure with targeted properties and performance. Compared to the current trial-and-error approach, mastering the Process-Structure-Property-Performance (Proc.→Struc.→Prop.→Perf.) relationships would speed up materials discovery with a huge societal impact (e.g. for energy transition). From a fundamental standpoint, there is no theory for these relationships. Thanks to highthroughput (HT) ab initio simulations, Struc.→Prop. (and hence Prop.→Perf.) can be well predicted and machine-learning (ML) approaches have been recently used as much faster surrogate models. But the simulation of the complete Proc.→Struc. is still out of reach, and ML approaches are hindered by the lack of data given the vast amount of possible process paths.
GEMPROMISE will establish a generative active learning approach to suggest process parameters leading to targeted properties, promoting a physical and chemical understanding of Proc.→Struc.→Prop.→Perf., as ultimate goal. Its key ideas emerged in a synergistic brainstorming between AYMONIER (experiments), RIGNANESE (simulations), and VANDERGHEYNST (ML): (i) a multimodal ML model will be developed to leverage experiments and simulations as direct and indirect data providers of varying quantity and quality, integrating these modalities through a joint latent space allowing for generation, (ii) a HT synthesis and characterization platform will be designed to close the loop and respond to the ML model queries, and (iii) a HT simulation framework will be devised for predicting Struc.→Prop. information to complement experiments.
To illustrate the concept, GEMPROMISE will give birth to a bottom-up, sustainable, and scalable method to produce new synthetic layered silicates with controllable band gaps. Once established, this approach can be extended to other processes, structures, properties, and hence applications.
This project has received funding from the European Research Council (ERC) under the European Union's Horizon Europe research and innovation programme under the grant agreement number 101224255.