Worldwide, industries face significant challenges that often require the discovery and design of new materials spanning sustainable energy, biotechnology or lightweight materials for transport applications — but materials development is slow and costly. For example, the process of drug discovery today can take an average of 12 to 15 years, with billions of dollars invested per drug and a 90 percent fallout rate.
The ability to predict the properties of materials from those of their molecular building blocks is therefore a grand challenge which impacts the biomedical, engineering, and chemical sectors and involves both big data analytics and model-driven strategies. Advances in computational power mean larger and more complex materials be simulated at the molecular scale. It is now possible to use high performance computers to calculate target properties of hundreds of thousand of materials, store them in databases and use them for prediction of novel materials. But materials yet to be discovered require powerful simulation tools which can predict properties in advance of their synthesis or under conditions such as high temperatures or pressures which are challenging to replicate experimentally.
Computer power alone is not enough to deliver on this challenge and fundamentally new strategies which offer improved predictive power without vastly increased computational cost are needed.
For more than five years, IBM Research has pioneered a new strategy for materials simulation, in collaboration with major research institutions, including the Science and Technology Facilities Council (STFC) Hartree Centre. This strategy has now been published in Reviews of Modern Physics, joining a select club of articles shaping the direction of physics as we know it today.
In the paper, we explain how assumptions currently employed to render simulation of large systems at the molecular scale tractable on modern computing systems can also erode predictive power and undermine confidence in the conclusions. We also review a novel strategy, electronic coarse graining, to address this fundamental problem. They show how the forces between molecules which drive their behaviour can be computed directly from a simplified representation of electronic responses without having to assume the “force laws” in advance. This step removes a major failure mode in conventional simulations – opening a new pathway toward powerful molecular models of complex materials.
So far, the strategy has been applied to only a small number of materials as test cases, but the success of these proof of concept demonstrators creates a platform for significant expansion. Already within reach as next steps are electronically coarse grained models for simple biological systems and rudimentary forms of self-assembly on the pathway to a fully realised next-generation framework for biomolecular simulation.
Combining improved physical models such as those built from electronic coarse graining methods with data-driven approaches will pave the way for more systematic and automated materials discovery.
In principle we imagine a future where not only can the properties of future materials be more reliably predicted at the molecular scale but, coupled with AI techniques, it should become possible for computing systems to suggest new synthesis and design concepts to accelerate and extend the discovery process.