April 18, 2017
Addressing uncertainty in atomistic machine learning
Electronic structure calculations have provided great insights into the fundamentals of designing materials, reactions, and devices. However, the computational demands of such calculations can be very intense. Our machine-learning package, Amp, can accelerate such calculations by orders of magnitude.
However, predictions are only useful if we know when they can be trusted. In a recent work, we address uncertainty in atomistic machine learning. We show the general types of errors that can arise, and outline strategies to avoid potential pitfalls. Interestingly, we show ways that the uncertainty can be isolated down to a per-atom level. With such approaches, we anticipate the usefulness of atomistic machine learning to increase, ultimately leading to network-enabled calculators that can provide electronic structure calculations at the speed of search.
You can read our new Perspective in Physical Chemistry Chemical Physics.