TECHNOLOGY AND INNOVATION CENTER

Selection of 30 Molecular Structures from 500,000 Candidates and Automated Experiments

Differences between traditional material R&D and new material R&D based on MI. Utilizing MI enables material development performance to be significantly enhanced.

As briefly mentioned in the first part of this interview, organic polymer materials (polymers) have complex structures, and new developments in this area have been slow due to a lack of integration between input and output data for design. Even with the same composition, changes in polymerization and sample preparation conditions can alter the structure, causing fluctuations in material properties. Additionally, compared to other materials, the simulation accuracy for polymers is lower, presenting another challenge.

To address this, we utilized MI for efficiently exploring material candidates by developing technology to generate high-quality, large-scale data and to predict material properties with high accuracy. To achieve this, we focused on developing technologies that combines automated molecular simulations with automated robotic experiments. This effort targeted polymer and composite materials, specifically focusing on substrate materials for high-frequency applications.

Daikin's vision for MI: development of automated molecular simulations and automated robotic experiments. Machine learning predicts material properties, facilitating rapid proposals of candidate materials and enhancing understanding of material design principles.

First, we conducted molecular simulations on a computer, input molecular structures, and automatically calculated the targeted properties. We also built a machine learning model that predicts material properties based on molecular structures. From there, the model explores and proposes promising molecular structure candidates that could achieve the targeted properties, such as low dielectric constant, low dielectric loss tangent, and low thermal expansion coefficient, to facilitate the development of new materials.

Specifically, we used machine learning to narrow down 500,000 polymer structures to 1,000 candidates, and through multi-stage screening using proxy indicators, we eventually identified 30 promising molecular structures. Traditional methods would have required an enormous amount of time to simulate and select these materials, but this approach efficiently narrowed down the candidates in about three months.

Approach for exploring new molecular structures using MI. Machine learning models and proxy indicators for multi-stage screening efficiently identify candidate structures with promising target properties.

We also developed a prototype for automated experiments using robotics in collaboration with a lab at the University of Tokyo. This system automates the entire process, from sample preparation to property measurement (such as dielectric constant), which was previously done manually. By integrating a conveyor belt for sample transfer and coordinating robotic arms with various devices, we can measure material properties with human-level accuracy and high reproducibility, while also accumulating the corresponding experimental data.

Flow of automated experiments. A prototype of an automated system was developed to streamline the entire process, from film formation using heat press to property measurement (dielectric constant).

The new materials identified in this project are promising candidates for use in the upcoming 5G and 6G communication technologies.

In-House Data Infrastructure Uses Private Cloud to Automate Collection and Processing of Experimental Data

—Another key pillar of your approach is development of the material data accumulation infrastructure. Could you provide us the background and describe the goals?

Tomota: So, the question is why a data infrastructure is necessary. No matter how advanced MI technology becomes, the importance of chemical experiments in the field remains unchanged. In recent years, there has been an increase in robot-assisted experiments, and managing data individually in formats like Excel or CSV format makes data-driven development difficult. To handle the growing amount of data, a powerful and flexible data infrastructure is essential.

Unfortunately, creating this data infrastructure in-house requires entirely different technologies from chemistry. This data infrastructure is specifically designed for material development and was built on a private cloud (AWS) that includes elements like databases, data pipelines, and data marts. Instead of processing experimental data locally on Excel, we integrated it with visualization and analysis tools, making it easy to output results and share them with the research team for accumulation and utilization.

Key requirements for the infrastructure included standardization from data collection to processing and visualization, flexibility to adapt to changes in data and tasks in the field, and prevention of tool proliferation and reliance on individual expertise for reviews. With these needs in mind, we designed a streamlined data processing flow.

 

Data pipeline concept. Commercial tools and services tend to be over-engineered, expensive, and time-consuming to learn. To create a foundation that is user-friendly for researchers, we built an in-house structure by pooling our internal expertise instead of outsourcing.

When it came to system development, we chose to create it internally since outsourcing would likely result in rework, and the system would become a black box. Fortunately, Daikin's Technology and Innovation Center (TIC) has a dedicated team specializing in data utilization and information technology. With a cloud environment in place for experimenting with new technologies and frequent sharing of best practices, TIC's structure and on-site technology experts proved to be a tremendous help. Within six months, chemical and IT engineers collaborated from a unified perspective to develop a data platform tailored for the niche field of chemical R&D.

— Could you tell us about the data pipeline—how it works and how it uses the cloud to efficiently and systematically collect and analyze experimental data?

The data pipeline refers to the process of gathering data, processing it into meaningful configurations, visualizing it, and then analyzing it. At present, dozens of such pipelines operate concurrently. Even for the same material, multiple pipelines are necessary to compare and examine various experimental conditions. The evaluation results are processed within minutes after researchers upload raw experimental data and measurement conditions and can be viewed and analyzed using BI tools.

A key developmental focus of these platforms was simplifying maintenance and operations. If the system is too labor-intensive, it becomes unsustainable. Similarly, high maintenance costs could jeopardize the project due to business constraints. Thus, reducing operational workload and costs was one of the essential requirements. By leveraging AWS, managing architecture through Infrastructure as Code (IaC), and automating workflows, both operational complexity and costs were reduced to less than one-tenth compared to commercial software.

Currently, this data platform is applied to a limited set of development themes. However, in the future, we plan to apply it to other themes and gradually expand it as needed. By providing a platform optimized for the needs of chemical R&D, we aim to support efficient and effective data-driven development.

 

Researchers aspiring to excel in MI can try polishing their skills at Daikin

— What is the future outlook for materials design, and do you have any messages for those interested in Daikin?