Integrating multiple materials science projects in a single neural network

Abstract

In data-intensive science, machine learning plays a critical role in processing big data. However, the potential of machine learning has been limited in the field of materials science because of the difficulty in treating complex real-world information as a digital language. Here, we propose to use graph-shaped databases with a common format to describe almost any materials science experimental data digitally, including chemical structures, processes, properties, and natural languages. The graphs can express real world's data with little information loss. In our approach, a single neural network treats the versatile materials science data collected from over ten projects, whereas traditional approaches require individual models to be prepared to process each individual database and property. The multitask learning of miscellaneous factors increases the prediction accuracy of parameters synergistically by acquiring broad knowledge in the field. The integration is beneficial for developing general prediction models and for solving inverse problems in materials science.

Designing a multilayer film via machine learning of scientific literature

“…Scientific knowledge is perpetuated through a series of human research activities, and past literature provides a roadmap to future discoveries 38 . Academic achievements are promising sources of information for MI 39 ; however, our survey of databases related to layered structures found that they are lacking in comprehensiveness. Moreover, multiple functionalities are implemented by taking into account trade-offs in some properties, not by solely relying on theory, and previous approaches have difficulty meeting such a complex target 40 – 42 .…”

Tackling the Challenge of a Huge Materials Science Search Space with Quantum‐Inspired Annealing

“…[10][11][12][13] However, the function is intrinsically hard to define because of the uniqueness of the solution problem. [1,11,13] Furthermore, the preparation of chemical structures itself is still a challenge even with cutting-edge deep learning techniques due to the current limitations of model complexity and computing power. [14] Other molecule generation techniques, such as Bayesian approaches [15] and deep reinforcement learning, [16] also tend to have similar limitations, such as excessive search space compared to the limited computing power.…”

“…[1,3,5] Various approaches have been proposed to solve the massive search space problem, such as defining an inverse function x ¼ f ML À1 (y), which predicts structures or experimental conditions from a target parameter y. [10][11][12][13] However, the function is intrinsically hard to define because of the uniqueness of the solution problem. [1,11,13] Furthermore, the preparation of chemical structures itself is still a challenge even with cutting-edge deep learning techniques due to the current limitations of model complexity and computing power.…”

Generative Models for Extrapolation Prediction in Materials Informatics

“…[1][2][3][4][5] Since the properties of materials are uniquely determined by the states of their constituent atoms, their observed structure-property relationships can be mimicked by machine learning models (Figure 1). [3][4][5] Their predictions are often more accurate than traditional theory-based predictions and simulations, especially in the cases of complex systems where the computational costs of the traditional approaches increase exponentially. [5][6][7] To describe the structures of materials, researchers typically consider their characteristic representations, such as molecular formulas and crystal structures.…”

“…[3][4][5] Their predictions are often more accurate than traditional theory-based predictions and simulations, especially in the cases of complex systems where the computational costs of the traditional approaches increase exponentially. [5][6][7] To describe the structures of materials, researchers typically consider their characteristic representations, such as molecular formulas and crystal structures. 1,3 In particular, numeric array-type expressions are frequently used in materials informatics because of their high computational processability.…”