Sodium-ion batteries (SIBs) emerge as a key option for next-generation electrochemical energy storage, thanks to abundant sodium resources, cost-effectiveness, and excellent low-temperature performance—making them well-suited for electric vehicles (EVs) and large-scale grid storage. However, their wider commercial use is limited by the low energy density of cathode materials, a critical bottleneck in narrowing the performance gap with lithium-ion batteries. Among promising cathodes, NASICON (sodium superionic conductor)-structured materials like Na₃V₂(PO₄)₃ draw attention for their stable 3D frameworks enabling efficient Na⁺ diffusion, but they suffer from low multi-electron reaction utilization, reliance on toxic vanadium, and inefficient development via traditional "trial-and-error" methods.
To tackle these challenges, a team led by Prof. Lei Li from Beijing Institute of Technology (BIT), with collaborators, developed a data-driven machine learning (ML) approach, leading to the creation of Na₃Mn₀.₅V₀.₅Ti₀.₅Zr₀.₅(PO₄)₃ (NMVTZP)—a new NASICON cathode setting a new SIB performance benchmark. Prof. Li noted that unclear links between atomic/crystalline configurations and energy density long hindered rational cathode design, while their ML framework addresses this by identifying key material descriptors, predicting high-performance candidates, and redefining development efficiency.
The team first built a comprehensive dataset of 73 data points from 51 published studies, focusing on NASICON cathodes with the general formula NaxMy(PO₄)₃ (M = transition metals). They trained and validated four ML models: attention-Bayesian neural networks (AttenBNN), random forest (RF), support vector machine (SVM), and K-nearest neighbor (KNN). Rigorous analysis identified three critical factors governing NASICON cathode energy density: higher entropy (activating multi-electron redox reactions like Mn²⁺/Mn³⁺/Mn⁴⁺ and V³⁺/V⁴⁺/V⁵⁺ to boost specific capacity), lower equivalent electronegativity (enhancing ionic bond stability and optimizing Na⁺ diffusion channels), and a lattice parameter c of 21.4–22.2 Å (balancing structural integrity and low Na⁺ diffusion barriers). Among models, RF achieved the highest prediction accuracy with a mean absolute error (MAE) of 0.083, while AttenBNN was key for quantifying prediction uncertainty—vital for reliable material screening with small datasets.
Guided by ML predictions, the team synthesized NMVTZP via a sol–gel method. The resulting NMVTZP cathode showed exceptional performance surpassing previous NASICON materials: it delivers a reversible specific capacity of 148.27 mAh g⁻¹ at 0.1 C, with an average operating voltage of 3.14 V, translating to an energy density of 465 Wh kg⁻¹—a significant improvement over traditional Na₃V₂(PO₄)₃ (396 Wh kg⁻¹). In rate capability, it retains 90.20 mAh g⁻¹ at 5 C and maintains stable discharge at 10 C, meeting fast-charging needs. Its cycling stability is also impressive, with 78.1% of initial capacity retained after 400 cycles at 5 C; in-situ X-ray diffraction (XRD) confirmed a mixed "solid-solution + biphasic" Na⁺ storage mechanism, ensuring structural stability during charge–discharge cycles. Additionally, NMVTZP addresses sustainability by replacing part of toxic vanadium with Mn, Ti, and Zr, and a 3.75 nm amorphous carbon coating further boosts electronic conductivity, laying a foundation for potential industrial scaling.
Looking forward, the team outlined future research directions to expand their work’s impact: extending the dataset to include other SIB cathode structures, integrating ensemble learning for multi-performance optimization, and applying the ML framework to electrolyte and anode development—aimed at accelerating full-cell SIB performance breakthroughs. They also plan to optimize NMVTZP’s large-scale synthesis through industry collaborations. Prof. Li emphasized that ML transforms the traditionally slow, costly material development process into a precise, efficient workflow, noting this ML-designed cathode is set to drive SIB commercialization in applications like photovoltaic storage and low-speed EVs, ultimately contributing to a more sustainable energy future.
Sodium-ion batteries (SIBs) emerge as a key option for next-generation electrochemical energy storage, thanks to abundant sodium resources, cost-effectiveness, and excellent low-temperature performance—making them well-suited for electric vehicles (EVs) and large-scale grid storage. However, their wider commercial use is limited by the low energy density of cathode materials, a critical bottleneck in narrowing the performance gap with lithium-ion batteries. Among promising cathodes, NASICON (sodium superionic conductor)-structured materials like Na₃V₂(PO₄)₃ draw attention for their stable 3D frameworks enabling efficient Na⁺ diffusion, but they suffer from low multi-electron reaction utilization, reliance on toxic vanadium, and inefficient development via traditional "trial-and-error" methods.
To tackle these challenges, a team led by Prof. Lei Li from Beijing Institute of Technology (BIT), with collaborators, developed a data-driven machine learning (ML) approach, leading to the creation of Na₃Mn₀.₅V₀.₅Ti₀.₅Zr₀.₅(PO₄)₃ (NMVTZP)—a new NASICON cathode setting a new SIB performance benchmark. Prof. Li noted that unclear links between atomic/crystalline configurations and energy density long hindered rational cathode design, while their ML framework addresses this by identifying key material descriptors, predicting high-performance candidates, and redefining development efficiency.
The team first built a comprehensive dataset of 73 data points from 51 published studies, focusing on NASICON cathodes with the general formula NaxMy(PO₄)₃ (M = transition metals). They trained and validated four ML models: attention-Bayesian neural networks (AttenBNN), random forest (RF), support vector machine (SVM), and K-nearest neighbor (KNN). Rigorous analysis identified three critical factors governing NASICON cathode energy density: higher entropy (activating multi-electron redox reactions like Mn²⁺/Mn³⁺/Mn⁴⁺ and V³⁺/V⁴⁺/V⁵⁺ to boost specific capacity), lower equivalent electronegativity (enhancing ionic bond stability and optimizing Na⁺ diffusion channels), and a lattice parameter c of 21.4–22.2 Å (balancing structural integrity and low Na⁺ diffusion barriers). Among models, RF achieved the highest prediction accuracy with a mean absolute error (MAE) of 0.083, while AttenBNN was key for quantifying prediction uncertainty—vital for reliable material screening with small datasets.
Guided by ML predictions, the team synthesized NMVTZP via a sol–gel method. The resulting NMVTZP cathode showed exceptional performance surpassing previous NASICON materials: it delivers a reversible specific capacity of 148.27 mAh g⁻¹ at 0.1 C, with an average operating voltage of 3.14 V, translating to an energy density of 465 Wh kg⁻¹—a significant improvement over traditional Na₃V₂(PO₄)₃ (396 Wh kg⁻¹). In rate capability, it retains 90.20 mAh g⁻¹ at 5 C and maintains stable discharge at 10 C, meeting fast-charging needs. Its cycling stability is also impressive, with 78.1% of initial capacity retained after 400 cycles at 5 C; in-situ X-ray diffraction (XRD) confirmed a mixed "solid-solution + biphasic" Na⁺ storage mechanism, ensuring structural stability during charge–discharge cycles. Additionally, NMVTZP addresses sustainability by replacing part of toxic vanadium with Mn, Ti, and Zr, and a 3.75 nm amorphous carbon coating further boosts electronic conductivity, laying a foundation for potential industrial scaling.
Looking forward, the team outlined future research directions to expand their work’s impact: extending the dataset to include other SIB cathode structures, integrating ensemble learning for multi-performance optimization, and applying the ML framework to electrolyte and anode development—aimed at accelerating full-cell SIB performance breakthroughs. They also plan to optimize NMVTZP’s large-scale synthesis through industry collaborations. Prof. Li emphasized that ML transforms the traditionally slow, costly material development process into a precise, efficient workflow, noting this ML-designed cathode is set to drive SIB commercialization in applications like photovoltaic storage and low-speed EVs, ultimately contributing to a more sustainable energy future.

The complete study is accessible via DOI: 10.34133/research.0794