Engineers at Cornell University have introduced a new artificial intelligence framework designed to predict and interpret the chemistry behind high-conductivity lithium-ion battery electrolytes. Led by Professor Fengqi You at the Cornell Duffield College of Engineering, the study presents a modeling platform that not only forecasts electrolyte performance but also reveals how salts, solvents and operating conditions interact to govern ion transport.
Wang, Z., & You, F. (2026). A dynamic routing-guided interpretable framework for salt–solvent chemistry. Nature Computational Science. https://doi.org/10.1038/s43588-026-00955-5
Electrolytes are the medium through which lithium ions move between electrodes during battery operation. In high-energy lithium-ion systems, nonaqueous liquid or gel electrolytes are commonly used because they support wider voltage windows and higher energy density. However, electrolyte formulation is complex. Conductivity depends on multiple coupled variables, including salt concentration, solvent composition, temperature and molecular structure. Small changes in one component can alter solvation structure, ion pairing and mobility.
Professor Fengqi You from Cornell University stated,
“These two studies address both an overarching framework and a specific modeling approach, reflecting our effort to connect strategy with practical implementation in AI-driven battery research.”
Traditional machine learning models applied to electrolyte research often treat these variables as a flat set of inputs. They are trained to correlate formulation parameters with conductivity or stability metrics. While these models can achieve strong predictive accuracy, they frequently operate as black boxes. Engineers receive a numerical prediction without a clear understanding of the underlying chemical logic.
The Cornell team sought to address that limitation. Postdoctoral researcher Zhilong Wang, first author of the study, worked with You to develop a dynamic routing-guided framework that treats salts, solvents and operating conditions as distinct but interacting contributors. Instead of funneling all descriptors into a single model layer, the system processes chemically meaningful information in parallel streams. The framework then adaptively integrates these streams to produce a prediction.
This architecture allows the model to retain interpretability. Engineers can examine how each component influences conductivity and how interactions between components shift across operating regimes. The goal is not only to identify high-performing formulations but to clarify why they perform well.
When applied to a large dataset of lithium-ion electrolyte experiments, the model reduced prediction error by more than 65 percent compared with several leading machine learning approaches. Importantly, it maintained accuracy across the full range of conductivity values, including rare high-conductivity formulations that are particularly relevant for next-generation batteries. Conventional models often struggle in these sparse data regions because they are biased toward more common, moderate-performance samples.
Beyond predictive performance, the study emphasizes integration with chemical principles. By structuring the model around domain knowledge rather than purely statistical correlations, the framework bridges computational learning and physical understanding. According to the authors, this approach increases confidence when extrapolating beyond the original dataset.
The work connects with broader efforts within Cornell’s AI4S Initiative, which aims to apply artificial intelligence to sustainability challenges in energy and materials. In parallel with the Nature Computational Science publication, You and Wang have also outlined a strategy for AI integration in solid-state battery research in a recent Science Advances article. That broader framework combines machine learning, simulations and experimental feedback loops.
In the case of liquid electrolytes, interpretability is particularly relevant. High conductivity often requires balancing ionic dissociation with solvent coordination. Too strong ion pairing reduces mobility, while too weak interactions may compromise stability. Temperature and concentration further complicate the landscape. An interpretable model can help clarify how formulation changes shift these competing effects.
From an engineering standpoint, the ability to rationally design electrolytes is central to improving energy density, charging speed and safety. As electric vehicles and grid storage systems scale, even incremental gains in electrolyte performance can translate into meaningful system-level improvements. AI tools that accelerate screening while preserving chemical insight may shorten development cycles.
The authors note that predictive accuracy alone is insufficient for industrial deployment. Engineers must understand failure modes and operating constraints. By structuring the AI framework around distinct chemical contributors, the model provides a clearer pathway for hypothesis generation and experimental validation.
The study reflects a broader shift in computational materials science. Rather than replacing physical reasoning, AI is increasingly being used to augment it. Interpretable architectures offer a way to combine data-driven efficiency with mechanistic transparency.
While further validation across expanded datasets and real-world battery systems will be necessary, the framework demonstrates that machine learning can move beyond correlation and toward explanation. For lithium-ion electrolytes, that shift may help engineers design formulations that achieve higher conductivity without sacrificing stability.
As battery research continues to integrate computation, experimentation and data science, tools like this dynamic routing framework illustrate how AI can be embedded within established engineering workflows. The result is not simply faster prediction, but a clearer view of the chemistry that drives performance.
Adrian graduated with a Masters Degree (1st Class Honours) in Chemical Engineering from Chester University along with Harris. His master’s research aimed to develop a standardadised clean water oxygenation transfer procedure to test bubble diffusers that are currently used in the wastewater industry commercial market. He has also undergone placments in both US and China primarely focused within the R&D department and is an associate member of the Institute of Chemical Engineers (IChemE).
