GEOMIND: A hybrid generative artificial intelligence for geopolymer design and optimization

Abstract

Geopolymers are an emerging class of eco-friendly materials with a wide range of applications. Nevertheless, achieving compounds for a specific application requires extensive experimental efforts on finding the accurate formulation of precursors. Using artificial intelligence, we tackle the challenge task of formulating accurate geopolymer mixtures that meets a predefined set of properties that the final materials should feature. This task, goes beyond the prediction of materials properties and focuses on the actual materials design. To this end, we build a high-quality in-house experimental database of geopolymer formulations and their physical properties. We develop a customly trained machine learning framework based on two variational autoencoder modules. The first, predicts the formulations that correspond to an array of target properties, and the second, corrects the requested properties to better match the predicted formulation. Furthermore, our model embeds a geopolymer feasibility bloc that ensures that the predicted materials can be synthetized. Overall, this framework is able to predict formulations and their corresponding properties with less than 10% error bar on a set of key properties of the final material encompassing the viscosity, the density and the compressive strength. The suggested methodology outperforms state of the art methods such as XGBR and Bayesian optimization, thereby paving the way for the systematic application of AI-based materials design in the development of eco-friendly novel materials for different applications.