Each encoder converts an input into a fixed sized vector of reduced dimensionality called latent representation. The decoder then reconstructs the input from the latent representation within a minimal reconstruction error. This procedure resembles that of zip files, where files (here, materials properties) are compressed into a single archive (here, a latent representation).
The latent representation forces the deep learning model to learn the important features in the data and minimize the noise. A neural network architecture links the features of latent space to the properties of the materials (Figure 2).
While public materials formulations are nearly non-existent, companies manufacturing materials make materials data part of their core digital business.
It is with such companies that we put our technology to the test. In close collaboration with Evonik, we applied our AI solution to the space of high-performance polymers. Together, we customized the entire AI architecture to predict polymer properties, refine existing formulations as well as generate new ones.
The AI models we designed address three types of predictions: material properties, compositions and processing parameters. Given a list of ingredients and quantities, the model can predict selected mechanical, physical and chemical properties of a product at specific test conditions, such as temperature or other parameters.
The model also takes into account selected process parameters that are part of a manufacturing process. The advantage here is that the predictions serve as a virtual experiment. Starting from ingredients and process parameters, the model predicts quantities that can be directly measured in the product at the end of the manufacturing process.
Essentially, the AI model predictions act as a compass, pointing researchers and chemists towards quicker breakthroughs in discovering new formulations for new materials to create innovative products.
In the best cases, our AI models proved to have an R-squared (R2) of 0.70-0.87 on unseen formulations. R2 is used as a metric of correlation calculating how close the true and the predicted values are. Values close to one indicate high accuracy predictions.
The same AI models can also predict recipes—ingredients and quantities—to produce materials with certain specifications. This should reduce the number of lab experiments by providing a set of predictions to start with.
Such AI architectures are general enough for a broad range of industrial problems. In addition to the work done with Evonik on polymeric materials, we’ve applied this algorithm in various material design processes, including in metallurgic industries for the design of metallic alloys and for the optimization of epoxy-resins. The potential goes beyond the scope of R&D with direct application in the production of materials, provided there are proper data curation processes for AI.
To explore all the possibilities of AI in depth, Evonik has expanded its cooperation with IBM in the field of digitization, and is the first chemical company to be part of the MIT-IBM Watson AI Lab.
Gaudin, T., Schilter, O., Zipoli, F., et al. Advanced Data-Driven Manufacturing. ERCIM News. 122 (2020). ↩