Data quantity governance for machine learning in materials science

Abstract

Data-driven machine learning is widely employed in the analysis of materials structure-activity relationship, performance optimization and materials design due to its superior ability to reveal latent data patterns and make accurate prediction. However, because of the laborious process of materials data acquisition, machine learning models encounter the issue of the mismatch between high dimension of feature space and small sample size (for traditional machine learning models) or the mismatch between model parameters and sample size (for deep learning models), usually resulting in terrible performance. Here, we review the efforts for tackling this issue via feature reduction, sample augmentation, and specific machine learning approaches and show that the balance between the number of samples and features or model parameters should attract great attention during data quantity governance. Following this, we propose a synergistic data quantity governance flow with incorporation of materials domain knowledge. After summarizing the approaches to incorporating materials domain knowledge into the process of machine learning, we provide examples of incorporating domain knowledge into governance schemes to demonstrate the advantages of the approach and applications. The work paves the way for obtaining the required high-quality data to accelerate the materials design and discovery based on machine learning.

Convolutional neural network prediction of the photocurrent–voltage curve directly from scanning electron microscopy images

“…However, the application of ML/DL for inorganic device development optimization remains challenging due to the limited availability of experimental data, especially at the research level, and this issue has been discussed in the literature. 7,8 High-throughput experiments that generate data automatically have also shown promise in material search and development as an experimental approach. [9][10][11][12] Another approach is to integrate the domain knowledge into the training of the model function to provide reasonable prediction.…”

“…[9][10][11][12] Another approach is to integrate the domain knowledge into the training of the model function to provide reasonable prediction. 8 The ML/DL techniques have recently been applied for photocatalytic materials and devices that have the ability of splitting water into oxygen and hydrogen using sunlight energy. 13 The application of ML/DL for inorganic device development optimization remains challenging due to the limited availability of experimental data.…”

“…13 The application of ML/DL for inorganic device development optimization remains challenging due to the limited availability of experimental data. 8 DL can be a powerful tool for analyzing data-rich environments, but when the available data are limited, ML approaches are more appropriate for optimizing chemical systems. We have studied hematite and bismuth vanadate (BiVO 4 ) electrodes, which absorb visible light and facilitate the evolution of oxygen, from a mechanistic viewpoint, 14,15 and for material optimization.…”

“…25 In this study, scanning electron microscope (SEM) images were used to predict the whole J-V curve using a convolutional neural network (CNN) for one of the typical and well-studied photocatalytic materials, BiVO 4 . 26,27 Although it is challenging that high-dimensional feature space of the CNN model is constructed with only dozens of sample images, as is described in the literature, 8 we could optimize the prediction network function by increasing the number of data by cutting image data and further optimization of the CNN network. Furthermore, we could extract dominant structural information for photocatalytic activity from the trained CNN function.…”

Parameter-Free and Electron Counting Satisfied Material Representation for Machine Learning Potential Energy and Force Fields

“…A configurational sampling relying on MD trajectories or energetic local minima from DFT calculations is highly sparse and may be insufficient to reflect the actual molecular motion. Recently, sample-oriented methods are frequently used to govern the data quantity . For example, the idea of active learning − is adopted, which performs an iterative process to gradually enlarge the data set size until convergence is achieved. , However, the data set may be highly sensitive to the initial sampling pool and remain biased and incomplete.…”