About us

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 high-throughput (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 Cyril AYMONIER (experiments), Gian-Marco RIGNANESE (simulations), and Pierre VANDERGHEYNST (ML):

1. 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,
2. a HT synthesis and characterization platform will be designed to close the loop and respond to the ML model queries, and
3. 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.