In a groundbreaking advancement from Chinese higher education institutions, researchers at Northeastern University have harnessed machine learning to design high-entropy ceramics based on barium titanate (BaTiO₃, often abbreviated as BT), achieving unprecedented energy storage performance. This innovation addresses long-standing challenges in developing lead-free dielectric capacitors capable of powering next-generation electronics, electric vehicles, and pulsed power systems.

The study, published in the Journal of Advanced Ceramics, showcases how artificial intelligence can navigate vast compositional spaces, slashing the time from years of trial-and-error to targeted experiments. Led by teams from Northeastern University's Shenyang and Qinhuangdao campuses, in collaboration with Shijiazhuang Tiedao University, this work exemplifies China's rising prowess in materials science research at its universities.

Understanding High-Entropy Ceramics and Their Role in Energy Storage

High-entropy ceramics represent a paradigm shift in materials design, where multiple principal elements are combined in near-equimolar ratios to maximize configurational entropy. This stabilizes single-phase structures with disordered lattices, enhancing properties like dielectric breakdown strength and polarization.

Dielectric capacitors store energy via electric polarization and release it rapidly, outperforming batteries in power density but lagging in energy density. Traditional BaTiO₃-based ferroelectrics suffer from hysteresis losses due to remanent polarization (P_r). High-entropy engineering disrupts long-range order, fostering relaxor behavior with slim polarization-electric field (P-E) loops, larger maximum polarization (P_max) minus P_r (ΔP), and thus higher recoverable energy density (W_rec = ∫(E dP from P_r to P_max)) and efficiency (η = W_rec/total energy input).

In China, where electric vehicle production leads globally—with over 9 million units in 2025—the demand for compact, high-power capacitors for regenerative braking, inverters, and fast charging is surging. Pulsed power applications in railguns, lasers, and grid stabilization also benefit from such materials.

The Machine Learning Framework Revolutionizing Materials Discovery

The Northeastern University team built a random forest regression model trained on 71 BaTiO₃-based bulk ceramics dataset, featuring compositions like (Bi_xNa_xBa_ySr_zCa_u)(Ti_aZr_bNb_c)O₃. This model predicted W_rec with high accuracy (R² ~0.9).

Coupled with the expected improvement (EI) acquisition function, it explored 660,000 candidates in the A-site and B-site high-entropy space, balancing exploitation of known good regions and exploration of novel ones. Top predictions guided synthesis of five Zr-doped variants: Ba_0.24Sr_0.24Bi_0.26Na_0.26Ti_1-xZr_xO₃ (x=0.12-0.18).

• Step 1: Data collection and feature engineering (ionic radii, tolerance factors, electronegativity).
• Step 2: Model training and validation via cross-validation.
• Step 3: Bayesian optimization-like screening using EI to select top 5 compositions.
• Step 4: Solid-state synthesis, sintering at 1250°C, and characterization.

This ML pipeline reduced experiments from thousands to mere dozens, accelerating discovery by orders of magnitude.

Record-Breaking Performance of the Optimal Composition

The standout ceramic, with x=0.15 (Ba_0.24Sr_0.24Bi_0.26Na_0.26Ti_0.85Zr_0.15O₃), delivered W_rec=10.8 J/cm³ at 600 kV/cm—63% above the training dataset max—and η=86%, with ΔP=51.02 μC/cm². This positions it among the highest for lead-free bulk ceramics under practical fields.

P-E loops were slim and linear, indicative of superparaelectric-relaxor ferroelectric crossover. Breakdown strength reached ~700 kV/cm, thanks to refined grains (~200 nm) and increased grain boundaries scattering charges.

Compared to prior BT-based high-entropy works (e.g., ~5-8 J/cm³ at higher fields), this achieves superior density at lower fields, ideal for devices.

Microstructure: The Secret to Superior Properties

X-ray diffraction (XRD) confirmed perovskite phase with orthorhombic-tetragonal-rhombohedral coexistence for x=0.15, promoting diffuse phase transitions. Scanning electron microscopy (SEM) revealed bimodal grains: fine ~200 nm matrix with larger ~1 μm inclusions, maximizing boundary density for enhanced insulation.

Transmission electron microscopy (TEM) disclosed nanodomains (~10 nm) and polar nanoclusters coexisting in the crossover region, flattening free energy landscape for reversible polarization.

Stability Under Real-World Conditions

The ceramic shone in thermal stability: W_rec ~5 J/cm³ (20-25% of max) and η >85% from 50-180°C, crucial for EV underhood use (up to 150°C). Frequency independence (10-100 Hz) and fatigue resistance over 10⁵ cycles underscore reliability.

Charge-discharge: At 400 kV/cm, discharge time t_0.9=40.7 ns, current density 1432 A/cm², power density 286 MW/cm³—ultrafast for pulsed apps.

Implications for China's Advanced Energy Technologies

China's dominance in EVs (BYD, NIO) demands capacitors for DC-link, traction inverters, and regenerative braking. High-density dielectrics enable lighter, smaller modules, boosting range 5-10%.High-power capacitors in EVs

In pulsed power—railguns (navy), high-energy lasers—nanosecond discharge supports megawatt pulses. Smart grids benefit from stable storage amid renewables intermittency.

This aligns with China's 14th Five-Year Plan for materials innovation, positioning universities like Northeastern as hubs.

Northeastern University's Leadership in Computational Materials Science

Northeastern University, a top-tier Chinese institution, hosts the Key Laboratory of Dielectric and Electrolyte Functional Materials. Prof. Xiaoyan Zhang's group integrates ML with experiments, following successes in relaxors. Collaborations expand interdisciplinary expertise.

China's universities lead ML-materials: Tsinghua, PKU publish heavily, with national funds like NSFC fueling 1000+ ML-papers yearly in ceramics.

Challenges Overcome and Future Directions

Past hurdles: combinatorial explosion in high-entropy (10¹¹⁺ spaces), synthesis uniformity, scalability. ML resolves via predictive screening; nanoscale control via doping/grain refinement.

• Challenges: Data scarcity (addressed by transfer learning potential), industrial scaling.
• Solutions: ML-active learning loops, high-throughput experiments.

Outlook: Target 15+ J/cm³, thin-film ML-design for >20 J/cm³, integration in EV prototypes. Broader: ML for superconductors, thermoelectrics at Chinese unis.

Read the full paper for details: Journal of Advanced Ceramics.

Photo by Max Berger on Unsplash

Stakeholder Perspectives and Global Impact

Researchers hail ML as 'game-changer': 'Drastically reduces cycle,' per Prof. Zhang. Industry eyes commercialization; policy supports via 'Made in China 2025'.

Global: U.S./EU lag in bulk ceramics; China's output 40% world ceramics research. Fosters international collab, e.g., with Japan high-entropy pioneers.