Predictive Materials Modeling

Alanine Dipeptide Ala2

The ability to predict the properties of a range of materials with confidence is paramount to a number of industrial sectors, including aerospace, pharmaceutical, electronics, automotive, and energy. These sectors are among those that will benefit from mathematical innovations that lead to the improved development and accelerated deployment of new materials.  

Computational materials models do not always perform as desired in the sense of being able to explain observation data under different process conditions. Discrepancies between observations and predictions may be generally explained by modeling uncertainties, phenomenological constitutive equations, or approximations and errors in electronic structure calculations, coupling of behaviors across scales from the continuum to the quantum scale, using surrogate models (e.g. cluster expansions in ab-initio simulations), material-specific interatomic potentials in molecular simulations, and many more.

Additional challenges in predictive multiscale materials models include the inability to model high-dimensionality problems, accounting for information loss when coarse-graining, modeling of rare events, lack of scalable uncertainty quantification techniques, and other issues.

The revolution that has taken place over recent years in the data sciences, along with the rapid advances in high performance computing and in the development of experimental means to collect structure and property data at different resolutions and scales, provide an opportunity for a data-driven predictive modeling and design of material systems.

This research brings together leading scientists in computational materials science, uncertainty quantification, machine learning communities, and industry leaders to address data-driven predictive multiscale materials modeling and design.

This fundamental work at the CICS on innovative mathematical, statistical, and machine learning techniques (e.g. development of surrogate models based on deep learning or on Generative Adversarial Networks) is of paramount importance for the accelerated design of materials and processes in the presence of inherent, structure, process, model, and epistemic uncertainties. In addition, active learning and experimental design techniques are used to identify the most informative experiments to improve predictive capabilities.  

The information-theoretic approach to materials modeling and design is based on an interdisciplinary research that has the potential for significant impact and relevance to technologically important industries.

Faculty involved in this research area include Karel MatousNicholas Zabaras, and others.