logo
ResearchBunny Logo
Generative learning facilitated discovery of high-entropy ceramic dielectrics for capacitive energy storage

Engineering and Technology

Generative learning facilitated discovery of high-entropy ceramic dielectrics for capacitive energy storage

W. Li, Z. Shen, et al.

Discover groundbreaking advancements in high-entropy dielectrics! A team of researchers, including Wei Li and Zhong-Hui Shen, achieved an impressive energy density of 156 J cm⁻³ at 5104 kV cm⁻¹, far surpassing previous benchmarks. This innovative approach not only accelerates material discovery but also paves the way for enhancing dielectric applications in electronics.

00:00
Playback language: English
Introduction
Dielectric capacitors, storing and releasing charges via electric polar dipoles, are essential in modern electronics due to their higher power density compared to electrochemical counterparts. However, their energy density (Ue) is significantly lower than chemical energy storage devices like batteries, limiting their applications, particularly in miniaturization and integration for the Internet of Things. Improving Ue is crucial. While ferroelectrics like BaTiO₃ (BTO), BiFeO₃ (BFO), and PbZrTiO₃ (PZT) offer high maximum polarization (Pm), their large residual polarization (Pr) and low breakdown strength (Eb) hinder high Ue. The high-entropy strategy, introducing local disorder, has emerged as a potential solution. High-entropy dielectrics (HEDs) with configurational entropy (Sconfig) ≥ 1.5R achieve a diverse polarization configuration, leading to smaller, weakly coupled polar nanoregions (PNRs), reducing Pr and increasing Eb due to increased lattice distortion. This synergistic effect significantly improves energy storage performance. However, the vast compositional space of HEDs makes traditional trial-and-error methods inefficient and time-consuming. Machine learning offers an alternative, but limited data can lead to issues like overfitting. This work introduces generative learning to address this data scarcity problem, creating a framework to accelerate the discovery of high-energy-density HEDs.
Literature Review
Existing research has highlighted the potential of high-entropy dielectrics for enhancing energy storage performance. Studies have demonstrated that the introduction of multiple cations and anions in a perovskite structure can lead to local structural disorder, resulting in a reduction in residual polarization and an increase in breakdown strength. Phase-field simulations have been used to model the impact of configurational entropy on the polarization response, and experimental work has validated the effectiveness of the high-entropy strategy in various dielectric systems. However, the vast compositional space and the limited data available for many high-entropy systems have hindered their wider adoption. Previous efforts have focused primarily on experimental trial and error, which is time-consuming and inefficient. This research attempts to address this challenge by utilizing machine learning techniques to guide the search for optimal high-entropy dielectric compositions.
Methodology
This study employed a three-part generative learning framework: (i) a generative model for generating latent space representations of high-entropy dielectric compositions, (ii) a classification and sampling stage for selecting promising candidates, and (iii) a forward inference and inverse design step for predicting and ranking compositions based on predicted energy density. Phase-field simulations were conducted to understand the impact of configurational entropy on polarization and energy storage properties in various perovskite systems (BiFeO3, BaTiO3, PbTiO3). The experimental part focused on Bi(Mg<sub>0.5</sub>Ti<sub>0.5</sub>)O₃ (BMT), a ferroelectric with relatively good stability. 77 experimental results of Bi₁₋ₐ₋в₋꜀LaₐSrբCa꜀(Mg₀.₅Ti₀.₅)₁₋d₋ₑ₋fMn<sub>d</sub>Zr<sub>e</sub>Hf<sub>f</sub>O₃ compositions were used as initial data to train the generative model, which employed an encoder-decoder architecture with an artificial neural network (ANN). This model was used to generate a large dataset of potential high-entropy compositions, enabling the exploration of a virtually infinite compositional space (nearly 10<sup>10</sup> combinations). A classifier was used to identify compositions with high energy density, which were then subjected to a ranking strategy based on predicted energy density and uncertainty, selecting the top five compositions for experimental validation. Films were fabricated via chemical solution deposition (CSD), and their structural and electrical properties were characterized using various techniques (GI-XRD, FE-SEM, HR-TEM, PFM, impedance analyzer, and ferroelectric workstation).
Key Findings
Phase-field simulations confirmed that increasing Sconfig leads to a more disordered dipole distribution, transitioning from ferroelectric to relaxor-like behavior, and significantly increasing Ue. Initial experimental trials using a trial-and-error approach yielded a maximum Ue of 87 J cm⁻³, only three times that of pristine BMT, highlighting the limitations of traditional methods. The generative learning model successfully generated 2144 sets of high-performance systems with energy densities exceeding 65 J cm⁻³. Five compositions were selected based on a ranking strategy that combined predicted energy density and uncertainty. Experimental validation of these five compositions revealed significant improvements in energy storage performance. Specifically, the Bi<sub>0.87</sub>La<sub>0.08</sub>Sr<sub>0.05</sub>Ti<sub>0.41</sub>Mg<sub>0.39</sub>Mn<sub>0.15</sub>Zr<sub>0.05</sub>O₃ film (C-3) exhibited the highest Ue of 156 J cm⁻³ at 5104 kV cm⁻¹, an eightfold increase compared to BMT. GI-XRD, SEM, and HR-TEM analyses indicated grain refinement and an increase in the proportion of amorphous phases in the high-entropy films, which may contribute to the improved energy storage properties. The high-entropy films also demonstrated lower dielectric loss, enhanced breakdown strength, reduced leakage current, and improved fatigue, temperature, and frequency stability.
Discussion
The findings demonstrate the effectiveness of integrating generative learning with experimental validation for accelerating the discovery of high-performance HEDs. The generative model successfully navigated the vast compositional space, identifying promising compositions with high energy density. The ranking strategy effectively managed uncertainty in the model predictions, leading to the successful identification of multiple high-performing HEDs. The significant improvement in energy density observed in the C-3 film underscores the potential of high-entropy design for enhancing energy storage performance. The synergistic effects of reduced Pr, increased Eb, and improved insulation contribute to the superior properties of the high-entropy films. The improved temperature and frequency stability further enhances their applicability in energy storage devices.
Conclusion
This study presents a novel and efficient methodology for designing high-entropy dielectrics with superior energy storage capabilities. By combining phase-field simulations, generative learning, and targeted experiments, the research significantly reduced the experimental burden associated with exploring the vast compositional space of HEDs. The successful synthesis of a high-entropy BMT-based film with an eightfold increase in energy density showcases the potential of this approach. Future research could focus on expanding the scope of this method to other material systems and exploring the underlying mechanisms responsible for the enhanced energy storage performance of HEDs. Further investigation into the relationship between local structural features and dielectric properties could provide valuable insights for designing even more advanced energy storage materials.
Limitations
The study focused on a specific class of high-entropy dielectrics based on BMT. While the generative learning model proved successful in this system, its generalizability to other material systems requires further investigation. The accuracy of the model's predictions depends on the quality and quantity of the initial experimental data. Although the ranking strategy mitigated some of the uncertainty associated with the model, there is still some degree of risk in selecting candidate compositions based on predictions alone. Future research should address these limitations to expand the applicability and robustness of this generative learning approach.
Listen, Learn & Level Up
Over 10,000 hours of research content in 25+ fields, available in 12+ languages.
No more digging through PDFs, just hit play and absorb the world's latest research in your language, on your time.
listen to research audio papers with researchbunny