In the quest for revolutionizing energy storage and electrochemical systems, the intricate dance between salts and solvents has once again taken center stage. This fundamental chemistry underlies not only the ionic conductivity essential to battery performance but also dictates the viscosity and chemical stability of electrolytes—key parameters for the efficiency and durability of devices ranging from lithium-ion batteries to supercapacitors. Yet, despite its critical role, advancing salt–solvent chemistry has been hampered by the enormous complexity inherent in the wide chemical space of solvent and salt combinations. This complexity is further exacerbated by nonlinear interactions and sparse, imbalanced experimental data that thwart conventional models from accurate generalization.
Enter SCAN, a novel, dynamic routing-guided framework poised to redefine how scientists model and interpret salt–solvent chemistry. Developed by researchers Wang and You, SCAN is designed to overcome the dual challenges of chemical complexity and data scarcity, leveraging advanced computational methods to unlock unprecedented predictive power and interpretability. This represents a significant leap in the field, as SCAN’s design permits it to intelligently navigate the labyrinthine chemical space while handling the often skewed datasets characteristic of experimental research in electrolytes.
At the heart of SCAN is its dynamic routing mechanism, inspired by breakthroughs in deep learning architectures. Unlike static models, SCAN dynamically routes information based on input data characteristics, allowing it to adaptively weigh features according to their relevance in diverse salt–solvent systems. This flexibility equips SCAN to capture the full spectrum of formulations, from well-studied to scarcely characterized combinations, making it a robust tool in the arsenal of electrochemical materials research.
The efficacy of SCAN was rigorously tested on non-aqueous electrolyte systems, a domain of particular interest due to their relevance in high-energy-density batteries. Achieving a benchmark mean absolute error of just 0.372 mS cm⁻¹ when predicting ionic conductivity, SCAN reduced the error margin by an impressive 65.3% compared to existing baseline models. This stellar performance not only signals improved predictive accuracy but also paves the way for accelerated discovery and optimization of electrolyte formulations.
What truly sets SCAN apart is not just its ability to predict but also its inherent interpretability, a rare trait in sophisticated machine learning models that often operate as black boxes. The framework integrates gradient-decoupling techniques—a mathematical approach that disentangles intertwined variable effects—symbolic regression, which generates human-readable equations derived from data patterns, and quantum chemistry calculations. Together, these methodologies offer a clear window into the mechanistic underpinnings of conductivity as influenced by molecular flexibility and ion–solvent interactions.
With SCAN, researchers were able to construct an expansive conductivity atlas encompassing over 11 million salt–solvent combinations, a feat previously unimaginable given the sheer size of this chemical space. This atlas serves as a comprehensive roadmap for identifying high-performance electrolyte candidates, significantly reducing the trial-and-error element historically associated with electrolyte design.
Experimental validations provide compelling evidence for SCAN’s practical utility. Across a massive candidate pool, the framework achieved an 81.08% success rate in identifying top-performing systems exhibiting conductivity values exceeding 20 mS cm⁻¹. Notably, this includes electrolytes based on prominent lithium salts such as LiFSI, LiTFSI, and LiBOB, which are widely regarded as promising candidates for next-generation battery technologies.
These findings hold transformative implications for battery research. By enabling precise tailoring of electrolyte properties, SCAN accelerates development cycles and optimizes performance attributes critical for enhanced energy density, longevity, and safety. Beyond batteries, the methodologies encapsulated in SCAN have potential applications in various electrochemical systems, including capacitors and fuel cells, where salt–solvent interactions dictate operational efficiency.
The researchers’ innovative use of symbolic regression, in particular, sheds light on quantitative structure–property relationships, distilling complex chemical behavior into interpretable mathematical expressions. This blend of explainability and performance is a significant stride toward bridging the gap between computational models and chemical intuition, fostering deeper insights into electrolyte chemistry.
Simultaneously, SCAN’s handling of imbalanced datasets addresses a long-standing challenge in materials science, where experimental data often skew toward popular or easily synthesized formulations. By dynamically adjusting to data distribution nuances, the framework maintains predictive accuracy across rare and underrepresented chemistries, expanding the horizon of potential discoveries.
Another hallmark of SCAN lies in its incorporation of quantum chemistry calculations to elucidate the influence of molecular flexibility on ionic conductivity. This quantum mechanical perspective complements statistical learning, capturing subtle electronic and structural factors that govern ion solvation and transport—effects that empirical models traditionally overlook.
The development of SCAN underscores a broader trend in materials research: the integration of artificial intelligence and fundamental chemistry to tackle complexity. It exemplifies how interdisciplinary approaches combining data science, theoretical chemistry, and materials engineering can unravel intricate scientific puzzles, accelerating innovation cycles.
Looking ahead, the adoption of SCAN-based strategies could catalyze a paradigm shift in electrolyte development workflows. Researchers will be empowered to explore vast chemical landscapes virtually before targeted synthesis and testing, markedly optimizing resource allocation and reducing experimental bottlenecks.
Moreover, the transparent nature of SCAN’s interpretive outputs ensures that domain experts retain control and insight over the optimization process, fostering a symbiotic relationship between algorithmic precision and human expertise. This duality is critical for advancing scientific understanding while harnessing the full power of computational tools.
In conclusion, SCAN represents a groundbreaking advancement in the modeling of salt–solvent chemistry. By mastering the dual complexities of chemical diversity and data imbalance, it offers a powerful, interpretable, and scalable approach to electrolyte design. Its successful demonstration on massive datasets and subsequent experimental validation heralds a new era of data-driven electrochemical innovation that promises to fast-track the journey toward safer, more efficient, and higher-performing energy storage systems.
Subject of Research: Salt–solvent chemistry in electrochemical systems, focusing on ionic conductivity, viscosity, and chemical stability in non-aqueous electrolytes.
Article Title: A dynamic routing-guided interpretable framework for salt–solvent chemistry.
Article References:
Wang, Z., You, F. A dynamic routing-guided interpretable framework for salt–solvent chemistry.
Nat Comput Sci (2026). https://doi.org/10.1038/s43588-026-00955-5
Image Credits: AI Generated
DOI: https://doi.org/10.1038/s43588-026-00955-5
