Abstract
We will demonstrate examples of applications from computational chemistry and materials informatics where "normal", i.e. similar to that typically observed in other applications, accuracy of machine learning (ML) (exemplified by correlation coefficients R between reference and predicted values on the order of 0.9 or even less) is sufficient (such as ML of materials properties) as well as examples of applications where high, with test set R values better than 0.9999 or even 0.99999, accuracy is categorically required. We will demonstrate how this on one hand makes shine nonlinear ML methods where otherwise a polynomial regression would suffice, but on the other hand presents challenges, not observed in more mundane applications, for machine learning with neuromorphic devices that are currently actively researched and that are subject to circuit noise and other circuit instabilities as well as to batch variability of the neuron activation function. We show by numeric simulations how these factors prevent obtaining high ML accuracy. On the example of machine learning of the kinetic energy functional for orbital-free DFT, we show that while restrictions on the shape of the activation function are not critical, circuit noise and instabilities can prevent obtaining the required ML accuracy.
