IBM has created an algorithm that can intelligently tell the computer which simulations are worth running and focus resources on them. The hope is to make it easier for scientists to do materials discovery by making computationally intensive simulation runs more efficient.

A team of researchers from IBM and the University of Liverpool teamed up to create an algorithm that cuts CPU hours (CPUh) by over 500,000 on expensive simulation campaigns. This allowed them to identify new materials for gas storage quickly. The team describes their work in a research paper, “Accelerating Computational Discovery of Porous Solids Through Improved Navigation of Energy Structure Function Maps,” published in Science Advances. This research is part of a family of algorithms called Bayesian optimization, which has been applied in various scenarios.

IBM researchers have developed a simple service called IBM Bayesian Optimization (IBO) that allows users to easily get the value from using these algorithms without themselves having to become experts in optimization.

The application of Bayesian optimization to materials discovery is maturing as it forms a key part in IBM’s Accelerated Discovery strategy. The method provides computational savings and allows for the user to either do something they could not afford before or do more with their present budget than ever before.

To apply this technique to a real materials discovery problem, the research team asked their algorithm to simultaneously hunt for materials with good gas storage properties that are easy enough to observe in the lab.

In this research paper, the team demonstrates how they achieved success by extending their algorithm called Batch Generalized Thompson Sampling. They developed it to deal with multiple objectives (in this case, gas storage property score and lattice energy), which is why they show that it can be applied to accelerate a new technique known as Energy Structure Function (ESF) Maps. These ESF maps are powerful for computational materials design but expensive enough not to try for routine use.

Despite all of its progress, the IBM research team still faces some obstacles. For example, it’s difficult to work with a variety of constraints and objectives without passing that complexity onto the user. Another obstacle is taking into account this data from multiple sources during an optimization process.