A simple denoising approach to exploit multi-fidelity data for machine learning materials properties

Abstract

Machine-learning models have recently encountered enormous success for predicting the properties of materials. These are often trained based on data that present various levels of accuracy, with typically much less high-than low-fidelity data. In order to extract as much information as possible from all available data, we here introduce an approach which aims to improve the quality of the data through denoising. We investigate the possibilities that it offers in the case of the prediction of the band gap relying on both limited experimental data and density-functional theory relying different exchange-correlation functionals (with an increasing amount of data as the accuracy of the functional decreases). We explore different ways to combine the data into training sequences and analyze the effect of the chosen denoiser. Finally, we analyze the effect of applying the denoising procedure several times until convergence. Our approach provides an improvement over existing methods to exploit multi-fidelity data.

Probabilistic Neural Data Fusion for Learning from an Arbitrary Number of Multi-Fidelity Data Sets

“…In an increasing number of applications in engineering and sciences analysts have simultaneous access to multiple sources of information. For instance, materials' properties can be estimated via multiple techniques such as (in decreasing order of cost and accuracy/fidelity) experiments, direct numerical simulations (DNS), a host of physics-based reduced order models (ROMs), or analytical methods [1][2][3]. In such applications, the overall cost of gathering information about the system of interest can be reduced via multi-fidelity (MF) modeling or data fusion where one leverages inexpensive low-fidelity (LF) sources to reduce the reliance on expensive high-fidelity (HF) data sources.…”

“…MFNets accommodate noisy data and are trained via gradient-based minimization of a nonlinear least squares objective. While MFNets can learn non-hierarchical relations between data sources, they: (1) rely on having prior knowledge on a set of latent variables that explain the relations between the sources, (2) assume each source can be surrogated via a linear subspace model, (3) are not probabilistic and also require regularization, (4) impose independence assumption among the data sources to derive the likelihood (i.e., the objective) function, and (5) rely on iterative approaches for finding the optimal graph structure.…”