Pavlo Dral and his colleagues propose learning how to generalise from one quantum computing method to another, in order to make predictions that go beyond the quality of the training data.
Machine learning potentials (MLP) are increasingly replacing too costly quantum chemistry calculations to simulate molecules and materials. It is generally accepted that their accuracy is capped by the method that produced their drive data; So far, only a readjustment on experimental measurements allowed this limit to be crossed.
Generalize through methods
In a perspective article, Pavlo O. Dral, Yi-Fan Hou, Yuxinxin Chen and their co-authors defend another path, without experimental data: to train models to generalize not only between configurations and chemical compounds, but also between quantum calculation methods. Such a model "transferable between methods" can then predict the result of a method he has never seen.
The authors present functional implementations, both in density functional theory and wave function methods, which are more accurate than each of the methods used for training. In their view, this approach should change the way data sets are built and chemical simulation models are evaluated.
Read the full article on the original source
Go to the source →