Integrating experimental and simulation data via preference learning

Abstract

Practical experimental values usually provide high-accuracy results while computational simulation values usually connect to abundant low-accuracy results. We present a model on the basis of preference learning to integrate multi-fidelity data for a better prediction performance by adding more low accuracy data as a supplement of insufficient experimental data. However, in general, experimental and simulation data cannot be directly compared, and this is difficult task. Gaussian process preference learning and Bayesian optimization constitute the framework. To evaluate the performance of integration, several datasets of bandgap in oxides and wavelength in molecules from different sources have been trained, predicted and evaluated by the ranking and optimization iteration results. Compared to the models trained with only experimental data preference, models adding simulation data preference information lead to an improvement of performance, which confirms our primary hypothesis. Moreover, we have explored and discussed about how the quality of simulation data influence the prediction performance.