The Slow Pace of Material Innovation

The journey from conceptualizing a new material to its widespread market adoption is often a protracted one. Consider the evolution of smartphone screens, which required over a decade of meticulous design and manufacturing refinement to achieve enhanced durability. Similarly, lithium-ion batteries, conceived in the 1970s, did not reach commercialization until the 1990s. This lengthy development cycle, typically spanning 10 to 20 years, is largely attributable to the traditional trial-and-error methodologies employed in material design and improvement. This pace is increasingly at odds with the market's demand for novel, high-performance, and sustainable materials.

The Data Deluge and the Needle in the Haystack

In an era of industrial digitalization, R&D and manufacturing processes generate vast quantities of data. However, the inherent complexity of this data often prevents its full exploitation. Material manufacturing data typically includes detailed recipes, process parameters, aging conditions, and material properties such as ductility or malleability. The high-dimensional nature of this data space makes it challenging to adequately sample and identify meaningful patterns, even for seasoned domain experts. This situation is akin to searching for a needle in a haystack, where the sheer volume of information obscures the critical insights needed for innovation.

AI as a Pattern Recognition Powerhouse

To overcome these challenges, IBM Research, in collaboration with Evonik Industries and the MIT-IBM Watson AI Lab, is harnessing the power of Artificial Intelligence (AI). Inspired by AI architectures successfully used in image processing, the team is applying deep learning models to material manufacturing data. These models are adept at recognizing important features and identifying subtle yet crucial correlations and patterns within complex datasets. The core of this approach involves encoding material data, including recipes and process parameters, into a compressed, lower-dimensional representation known as a 'latent representation'. This process is analogous to data compression, where essential information is retained while noise is minimized. A neural network then links these latent features to specific material properties, enabling predictive capabilities.

Virtual Experimentation and Accelerated Discovery

The AI models developed can perform two critical functions: predicting material properties based on given ingredients and process parameters, and conversely, suggesting optimal recipes and process parameters to achieve a desired set of material specifications. This capability transforms the R&D process by enabling virtual experimentation, significantly reducing the reliance on time-consuming and costly physical laboratory tests. In the best-case scenarios, these AI models have demonstrated an R-squared (R2) value of 0.70-0.87 on unseen formulations, indicating a high degree of prediction accuracy. This predictive power acts as a compass, guiding researchers and chemists toward quicker breakthroughs and the development of innovative products.

Broad Industrial Applicability

The AI architectures employed in this collaboration are designed for broad applicability across various industrial challenges. Beyond the work with Evonik on polymeric materials, similar algorithms have been applied to material design processes in the metallurgic industries for the development of metallic alloys and for the optimization of epoxy resins. The potential extends beyond R&D, with direct applications in material production, provided that robust data curation processes are in place to support AI integration. This signifies a paradigm shift in how new materials are conceived, developed, and brought to market, promising a future of faster innovation and enhanced material performance across diverse sectors.

Tag: material design

IBM and Evonik Forge Ahead in Material Design with AI Collaboration