Electrostatic capacitors are essential energy storage devices across various sectors, including defense, aerospace, and transportation, offering superior power density compared to batteries or supercapacitors. However, a significant challenge lies in substantially increasing their energy density (Ue), particularly at high temperatures, to reduce size and weight. Current high-power capacitors use biaxially oriented polypropylene (BOPP), which performs well at room temperature but degrades rapidly at elevated temperatures. While alternatives exist with high thermal stability, they often compromise energy density due to lower band gaps (Eg). This limitation stems from the intricate relationship between a polymer's underlying chemistry and its function. The vast number of possible chemical variations in even a single polymer family makes exhaustive experimental exploration impractical. This study leverages artificial intelligence (AI) to navigate this expansive chemical space, enabling the rapid discovery of high-performance dielectrics. Over the past decade, AI has guided the discovery and characterization of materials, including polymer dielectrics with high energy density up to 100°C. Efficient materials discovery necessitates a strategy for selecting or generating a chemical subspace, estimating material properties within it, and selecting candidates for synthesis and testing. This research introduces the polyVERSE paradigm, which addresses these challenges by combining an expert system for polymer generation from commercially available monomers with multitask graph neural networks for property prediction. This AI-driven approach efficiently generates and screens a vast number of potential polymer structures, significantly accelerating the discovery process.

The literature review highlights the existing limitations of current high-temperature dielectric materials. Commercial polymers like BOPP, while offering low dielectric loss and a large Eg, suffer from low dielectric constant (ε) and poor high-temperature mechanical stability. Other high-thermal-stability polymers such as polyimides (PI), polyethertherketone (PEEK), polyetherimide (PEI), and fluorene polyester (FPE) trade stability for low Eg and consequently low Ue. Previous AI-driven approaches have shown promise in designing polymer dielectrics for high energy density up to 100°C but have not addressed the challenge of high-temperature performance beyond this range. The authors cite several previous works focusing on high-temperature energy storage materials, rational co-design approaches for polymers, and the application of machine learning to materials discovery. These works serve as a foundation for this research, which aims to push the boundaries of high-temperature dielectric performance.

The study employs the polyVERSE paradigm, a three-step AI-assisted materials discovery process. First, polymer structures are generated using an expert system that leverages established chemical reactions and commercially available monomers. This ensures synthetic feasibility by explicitly encoding reaction chemistry, unlike many previous methods. The focus initially centers on ring-opening metathesis polymerization (ROMP), a class known for high-temperature performance. A ROMP template, including a chemical transformation and a monomer filter, selects cyclic olefins based on criteria such as ring size and the absence of strong electrophilic groups. The second step involves property prediction using multitask graph neural networks (polyGNN). These networks learn the relationship between a polymer's chemical structure (represented as a graph) and its properties (Tg, Eg, and ε). The polyGNN algorithm's efficiency is enhanced by focusing on repeat units and leveraging GPU computation. The models are trained on large datasets, achieving good accuracy in predicting Tg, Eg, and ε. The third step involves screening. The generated polymers are screened based on predicted properties (Tg > 100°C, ε > 3, Eg > 4 eV), prioritizing structures with high Tg × ε × Eg. Five of the top candidates were selected for synthesis and experimental characterization. Synthesis was performed using ROMP, and the resulting films were characterized using various techniques including NMR, DSC, TGA, UV-Vis spectroscopy, and dielectric spectroscopy to determine Tg, Eg, ε, and breakdown field (Ebd). Later, a similar approach is used to explore polyimides, modifying the reaction template and screening criteria to identify candidates soluble in green solvents (water or ethanol). The polyGNN was also trained to predict polymer solubility in various solvents.

The research successfully identified a novel polynorbornene dielectric, PONB-2Me5Cl, exhibiting exceptionally high energy density over a broad temperature range. At 200°C, it achieved an energy density of 8.3 J cm⁻³, exceeding any commercial polymer by more than an order of magnitude. Below 200°C, PONB-2Me5Cl also surpasses all commercial polymers in energy density. The high performance is attributed to a combination of high Tg, large Eg, and moderate ε. The high Eg acts as a significant barrier to electron conduction, leading to an unprecedented breakdown field strength (Ebd). Experimental measurements closely match the AI predictions, validating the model's accuracy. The study also reveals the impact of subtle chemical changes on dielectric properties, highlighting the potential for further optimization. For example, modifying the backbone chemistry or aryl substituents significantly affects Ebd and U. Furthermore, the research explores the synthesis of polyimides soluble in green solvents (water or ethanol). Hundreds of polyimide structures meeting specific criteria for high-temperature dielectric properties and green solvent solubility were identified, opening up opportunities for environmentally friendly high-performance dielectrics.

The findings directly address the research question by identifying high-performance dielectrics exceeding the capabilities of commercially available materials. The exceptionally high energy density of PONB-2Me5Cl at high temperatures, along with its high breakdown strength and suitable dielectric constant, offers a significant advancement for energy storage applications. The success in discovering this material validates the effectiveness of the polyVERSE paradigm for accelerating materials discovery. The close agreement between predicted and experimental properties underscores the accuracy and reliability of the AI models used. The subtle variations in dielectric properties observed with minor chemical adjustments highlight the potential for fine-tuning material properties using AI-guided design. The exploration of polyimide structures soluble in green solvents points toward a more sustainable approach to manufacturing high-performance dielectrics. This research has implications for various sectors, particularly those requiring high-energy-density capacitors operating at high temperatures.

This study demonstrates the power of AI in accelerating materials discovery and design for advanced energy storage applications. The successful identification of PONB-2Me5Cl, a high-temperature dielectric with exceptionally high energy density, showcases the efficacy of the polyVERSE paradigm. The research also explores the potential of polyimides as a sustainable alternative for high-temperature applications. Future work could focus on optimizing the identified structures further, exploring other polymerization techniques, and investigating the incorporation of nanofillers to enhance performance. Expanding the polyVERSE database to include a wider range of monomers and reactions will further broaden the scope of material discovery.

The study focuses on a limited number of polymerization templates (ROMP and polyimides). While the chosen templates represent classes of polymers known for high-temperature stability, other polymerization techniques could lead to the discovery of even better-performing materials. The AI models rely on existing datasets and might not capture all the complexities of real-world material behavior, thus requiring experimental validation. The assessment of environmental impact focuses on specific aspects, such as lightweighting and solvent choice, and a more comprehensive life-cycle assessment would provide a more complete picture of sustainability.