Top ten methodologies in Materials Informatics and examples for how they can be used solve specific problems

The analysis and design of materials is often a slow process that takes weeks, months or years. And, many current material platforms rely on expensive raw material sources and fail to achieve sustainability goals. Meanwhile, Material Informatics – fueled by emerging techniques such as multiscale modeling, machine learning?and autonomous experimentation?– is transforming the way materials are discovered, understood, developed, selected, and can be used to develop materials with fewer resources and with greater impact.?What’s also exciting is that these methods can leverage legacy data and models we may already have available, but which have not yet been exploited.

Here are some key methodologies in Materials Informatics and how they can be used solve specific challenges:

1. Materiomics: Using analytical, simulation, and data-driven methods to break down materials into their genomic building blocks and then predict complex behaviors. E.g., we can discover the design principles of biological materials like silk, sea shells or bone and translate them into designer engineering materials made from ceramics, metals and polymers that have superior properties. An important frontier is the construction of findable, accessible, interoperable and reusable ("FAIR") data infrastructures, as done in support of the Materials Genome Initiative.
2. Designer Materials:?Solve inverse design problems to develop material recipes to meet a set of target demands with competing properties. E.g., we can use deep learning models to articulate materials with high strength and high toughness, and with tunable properties to change features during use, and generate products that use less resources while maximizing engineering properties and resilience (e.g., design lighter, more resilient battery materials with higher capacity at lower cost).?
3. Physics-based Deep Learning: Integrate physics models into deep learning frameworks for use in efficient simulations. E.g., we will learn how to predict materials failure using machine learning orders of magnitudes faster than using legacy modeling while incorporating known models with measured data to form a new class of predictive models, or we can amalgamate purely data-driven methods with solution approaches based on differential equations
4. Integrating Data Modalities:?Using transformer models, we learn how to integrate the use of image, video, text and graph data.?E.g., we can leverage legacy datasets for insights and apply it to design better sustainable materials made from inexpensive biocompatible resources
5. Computer Vision for Materials Engineering:?Integrate computer vision methods, graphic rendering, virtual reality, as well as interpretable machine learning to mine datasets. E.g., we can predict complex physical fields (e.g., stress, strain, temperature, phonons, scattering data, etc.)?from scarce sources and fill in missing parts in measurements or simulation that would otherwise be exceedingly expensive to calculate.
6. Large Language Models for Intelligent Materials Analysis: Applications in molecular modeling, microstructure generation, protein folding, including mining large datasets or published literature for human-readable design approaches, or to extract material construction, processing and design principles
7. Autonomous Materials Discovery: Integrating emerging learning strategies such as active learning, reinforcement learning with automated experimentation to autonomously explore vast design spaces. E.g., autonomous laboratories powered by AI tools proceed to explore, exploit and discover complex new materials with unseen functions, e.g. as applied in the discovery of bio-composites made from waste or unusual chemical compounds
8. Transfer learning and Fine-tuning: Techniques to use existing models or models trained for other problem spaces and adapt them to solve your specific problem. E.g. we will use pretrained computer vision models and develop microstructure analysis tools to assess failure risk using very small datasets?
9. Model Interpretability: Learning mechanistic insights from trained deep neural network, to inform engineering design, new scientific discovery and transfer solutions to problems
10. Integrated learning strategies: The use of reinforcement learning, active learning, and autonomous discovery offers important advantages over supervised methods, and can be combined effectively with experimental and characterization toolkits, including bio-inspired strategies of directed evolution and genetic adaptation of solutions

Scientific machine learning is a powerful platform technology

The broad methodological areas listed above can be broken down into a set of elementary machine learning methods that are the basis for solving Materials Informatics problems.

Some recent trends include:

• Computer vision methods, including convolutional neural nets and modern realizations such as ResNet, for tasks such as image regression and prediction of material properties from microstructures, image generation and field predictions, as well as segmentation problems
• Adversarial methods for generative tasks and cross-domain translations
• Geometric deep learning using graph neural networks, for analyzing, predicting and designing molecules, truss structures and other complex engineering designs (these include equivariant approaches that preserve geometric or physical invariants)
• Autoencoder architectures, for coarse-graining and the discovery of physical laws in complex applications
• Transformer architectures based on attention mechanisms, for solving multi-modality tasks at the interface of symbolic modeling, language, image/voxel data, and generative tasks to learn materiomic graphs to capture hierarchical structure-process-property relationships
• Diffusion models, for solving complex forward and inverse problems, including conditional generative design tasks for materials-by-design applications
• Efficient neural network architectures and hardware advances, to facilitate low-energy, efficient and resilient computations that reduce cost and environmental impact of computing

Other dimensions

There are many other dimensions to this emerging field, with significant opportunities ahead. These include, for instance:

• Multi-modality, such as the methods that allow us to conduct an integrated analysis of images, voxel data, dynamical data, and graphs,
• Language and symbolic methods and hybrid approaches
• Human-AI collaboration including AI reasoning and human interpretation, ethics, and biases
• The development, maintenance and use of materiomic databases
• Synthetic datasets, and data collection in materials development
• Visualization and data analysis methods, including statistical methods, cluster analysis, graphic rendering, virtual reality; as well as interpretable machine learning