This paper presents the application of materials informatics (MI) technology in the field of plastic recycling. In the field of plastic recycling, which is becoming increasingly important as we work toward the realization of a sustainable society, NEC has proposed a system for improving the efficiency of the plastic recycling process using MI technology and has confirmed the effectiveness of this system with a plastic recycler. We aim to bring about innovation in the field of plastic recycling.

1. Introduction

Efforts to reduce plastic waste, a major contributor to environmental pollution and habitat destruction, are gaining momentum worldwide. In Japan, national policies emphasize the 3Rs of plastics—Reduce, Reuse, and Recycle—together with proper disposal methods. Despite these initiatives, a substantial portion of plastic waste still ends up being incinerated or landfilled. Therefore, accelerating the reuse of waste plastics in the production of new products is vital for achieving a sustainable society.

Materials informatics (MI) has rapidly advanced as a digital approach utilizing information sciences like artificial intelligence (AI) to enhance material development. Applying MI effectively in the plastic recycling sector holds the promise of significantly improving the efficiency of reusing waste plastics and expanding their applications. However, the unique technical challenges associated with plastic recycling have limited the widespread adoption of MI in this area.

This paper presents the development and validation of an MI system specifically designed to address the challenges encountered in plastic recycling.

2. Challenges Faced by Plastic Recyclers

Plastic recyclers handle the process of transforming diverse waste plastics collected from factories and households into recycled plastic materials, a process known as material recycling. These recycled materials are then sold in the form of plastic pellets (referred to as recycled pellets). The production involves several stages: collecting and sorting waste plastics, crushing, compounding (which includes adjusting properties and color matching), and granulation. Despite this seemingly straightforward process, recyclers encounter several challenges at each stage, as outlined in the Table below.

Table Types of data obtained in the recycling process, challenges, and solutions.

In the production of recycled pellets, the compounding process is vital as it largely determines the final product's value. Currently, this process relies heavily on the experience and intuition of skilled workers, making it somewhat subjective and inefficient. Recyclers must carefully tailor the mix of various small-quantity, diverse waste plastics to meet customer specifications for attributes like strength, thermal fluidity, and color.

Additionally, many recyclers are small to medium-sized enterprises with limited financial resources, hindering their ability to invest in advanced machinery and internal systems.¹⁾ As a result, they often rely on paper-based workflows for managing data and inventory, which slows their progress towards digital transformation (DX).

To tackle these challenges, NEC has developed and validated a system by integrating its core MI technology with resin material technology. This system was created in collaboration with Maruki Sangyo Co., Ltd., a leading plastic recycler in Japan. It is designed to propose compositions, including physical property adjustments and color matching, that align with the dynamic demand and inventory status of waste plastics (Fig. 1).

This system enables recyclers to produce high-quality recycled pellets more efficiently. It also enables less experienced workers to perform at the level of seasoned experts, thus helping bridge the gap in specialized knowledge and expertise. This paper outlines the development of NEC's materials informatics (MI) system for recycled pellet production and presents the results of its successful implementation at Maruki Sangyo.

3. Transforming Recycled Pellet Production With MI Technology

The process of evaluating and determining the optimal composition for recycled pellets, including property adjustments and color matching, poses unique challenges compared to virgin materials. A significant issue is the lack of detailed information about the composition of recycled raw materials. Key data, such as molecular structure and physical properties readily available for virgin materials, are often missing in recyclables. Although it is possible to obtain this information using experimental equipment, the financial constraints and production speed demands faced by recyclers make this approach impractical.

To overcome these challenges, we investigated efficient methods for collecting the critical data necessary for materials informatics (MI)-based learning. This involved gathering input data such as the physical properties of waste plastics, color variations, compounding ratios, and additive details, as well as output data like predicted physical properties and color values of the recycled pellets. Based on these efforts, we developed a proprietary MI system to predict material properties and optimize color.

3.1 Development of the MI-based Physical Property Prediction System

NEC has developed an advanced materials informatics (MI) system to predict the physical properties of recycled pellets. This system is trained to assess how various types of waste plastics influence the production of pellets with specific properties. Due to incomplete analysis of incoming waste plastics, the system classifies materials based on their type—such as Acrylonitrile Butadiene Styrene (ABS) and Polystyrene (PS)—and inferred physical characteristics like strength and thermal fluidity, derived from product grade names. Representative values from these classifications are then used for MI learning. This method boosts the adaptability and broader applicability of the MI technology.

Given the limited availability of training data, the system employs linear and decision tree models integrated with feature engineering techniques from both physical and chemical perspectives to enhance prediction accuracy. Feature engineering focuses on essential factors for product design, including macroscopic material properties such as elastic modulus and specific gravity. These properties are generally derived by statistically analyzing microscopic physical and chemical characteristics. However, this innovative system estimates microscopic properties from macroscopic characteristics using representative values, which facilitates accurate linear calculations and the integration of polymer chain information, thus managing complex properties that were challenging to predict.

The MI-based system functions as a physical property database, enabling users to search for material properties by manufacturer and product name, significantly reducing manual work. This database allows users to explore representative values and variability ranges of physical properties for intermediate materials derived from recycled plastics, aiding marketing and sales efforts. From this perspective, the system is poised to become a critical component of inventory management for plastic recyclers, helping them effectively manage stock levels and ensure timely use of appropriate materials.

Looking ahead, the system is designed to generate physical property data that seamlessly integrates with supply chain management (SCM) and digital product passport (DPP) systems for inventory management. To facilitate this integration, data fields have been clearly defined and standardized. By establishing a framework for efficiently collecting high-quality data, the system is set to become an indispensable tool for recyclers in the future.

3.2 Development of the MI-based Color Matching Support System

The MI-based color matching support system predicts the color outcomes of recycled pellets based on the blend of multiple colorants. The mixing of colors follows two primary models: additive mixing (for light sources) and subtractive mixing (for materials like recycled pellets, which use dyes and pigments for coloring). There are two types of color: "light source color," which is the inherent color of light, and "object color," which results from light interacting with a material through reflection, absorption, or transmission. Object colors are expressed through the mixing ratios of the three primary subtractive colors—cyan (C), magenta (M), and yellow (Y). Additionally, mixing colors can either increase brightness (additive mixing) or decrease brightness (subtractive mixing), with materials such as paints, inks, and filters typically represented by the subtractive color model.

Color perception is determined by the absorption and reflection properties of pigments and dyes. Traditional computer color matching (CCM) systems store pigment data in spectral space and compute reflection and absorption spectra for each wavelength. In contrast, NEC has developed a data-driven MI-based color formulation system using the CIELAB color space (also known as Lab*). In this space, L* denotes lightness, a* represents the red-green axis, and b* represents the blue-yellow axis. Standardized by the International Commission on Illumination (CIE) in 1976, this color space accounts for material-dependent variations, including the inherent differences between dyes and pigments, making it particularly suited for use in recycled materials. Additionally, the system considers the impact of colorant concentration on final color expression, integrating these variations as key features in its learning model. This innovative approach enables a more efficient method for managing colorant data across different manufacturers.

Unlike the MI-based physical property prediction system, which requires a large dataset due to the complexity of recycled pellet material properties, the MI-based color formulation system uses fewer colorants. Because colorant data can be reused once measured, NEC conducted comprehensive spectral measurements for all colorants. For these measurements, high-transparency acrylic resin was selected as the base virgin material. Color reference plates were produced by adding 0.4 parts per hundred resins (phr) of pigment and 0.1 phr of dye. These plates were then subjected to reflectance measurements with a calibrated white background and transmittance measurements. The brighter measurement results (with higher L* values) were selected as reference data and used for model training.

4. Verification with Plastic Recycler

This section outlines the verification process and the results obtained from testing the two MI-based systems, conducted in collaboration with Maruki Sangyo.

4.1 Verification of the MI-based Physical Property Prediction System

During the verification process of the MI-based physical property prediction system, efforts were concentrated on enhancing its scalability. This was achieved by simultaneously conducting data organization and system development. For data organization, workflow interviews were held to gain insights into daily operations, extracting crucial data points relevant to both the inventory management system and the MI system requirements.

The dataset included 901 types of waste plastic materials, which were categorized into nine practical material groups based on material type and similar grade classifications, following the methodology described in section 3.1. Representative values for each group were measured. Material property values, serving as input data, were generated using Gaussian distribution with mean values and standard deviations for each category. Additionally, 69 types of additives used in recycled pellet production were classified into 10 groups. Their effects on material properties were visualized through graphs and modeled using quadratic curves based on concentration levels.

Through this comprehensive approach, NEC developed an MI-based physical property prediction system that balances versatility with reproducibility. Cross-validation (k=10) tests for specific gravity and impact strength showed relative errors of 3.4% and 30%, respectively, which are considered sufficiently accurate for practical applications. However, the melt flow rate (MFR), which had shown promising accuracy during preliminary testing with a smaller dataset, experienced a decline in precision. This was attributed to overly broad material group classifications and the absence of correlation modeling between different material properties. Future improvements in classification accuracy and an expansion of testing parameters are anticipated to enhance prediction accuracy further.

A comparison of predictions made by Maruki Sangyo's skilled workers and those produced by the MI system revealed a 10% deviation, indicating that even the current classification approach yields meaningful results. In anticipation of future integration with a digital product passport (DPP) system, further research focused on improving data update speed, refining classification management, and developing rapid material property measurement methods.

4.2 Verification of the MI-based Color Matching System

In the cross-validation (k=10) of the MI-based color matching system, outlined in section 3.2, the ΔEab color difference was 1.93 with a linear model and 0.689 with AdaBoost. Both values are well below the acceptable tolerance of ΔEab=2.0 for delivery, demonstrating considerable practical accuracy. For usability evaluation, NEC conducted interviews with employees at Maruki Sangyo, who are the intended primary users of the system. Based on their feedback, NEC developed a graphical user interface (GUI) that displays color difference (ΔE) graphs and provides visualized reference colors corresponding to L*, a*, and b* values, as shown in Fig. 2.

Through this verification, NEC confirmed the effectiveness of both the MI-based physical property prediction system and the MI-based color matching system in a recycled pellet manufacturing environment. These systems are poised to address challenges in recycled pellet production by enhancing both quality and operational efficiency. Moving forward, NEC will continue its efforts in data collection and technology development to advance and implement more sophisticated MI systems for industrial applications.

5. Future Outlook

Following the enactment of the Act on Promotion of Resource Circulation for Plastics in 2022, all stakeholders—including businesses, consumers, as well as national and local governments—are required to actively engage in promoting plastic resource circulation.²⁾ As a result, demand for recycled plastics is anticipated to rise significantly in the coming years. To support the growth of the recycled plastics market, NEC plans to broadly implement its MI-based systems across the recycling industry, establishing a framework that ensures the stable supply of high-quality recycled plastics.

At the same time, NEC aims to enhance collaboration among recyclers, using these MI systems as a basis for data sharing. By integrating data from multiple recyclers, NEC can improve system accuracy, thereby increasing the value of recycled plastics and contributing to the growth and revitalization of the recycling industry. Through the advancement of materials informatics (MI) technology, NEC is committed to supporting the development of a circular economy for plastics, striving toward a future where recycled plastics are routinely used in product manufacturing.

References

Authors’ Profiles

Related URL: