# Leveraging Machine Learning for Screening Metal-Organic Frameworks with Selective CO2 Recognition for Early Thermal Runaway in Lithium-Ion Batteries

## TL;DR

This paper uses machine learning to find materials that can detect CO2 early in lithium-ion battery failures, improving battery safety.

## Contribution

A novel integration of Monte Carlo simulations and Random Forest models to screen MOFs for CO2 selectivity in battery safety applications.

## Key findings

- The RF model achieved high predictive accuracy (R2 > 0.92) in screening MOFs for CO2 recognition.
- Key performance drivers include Q0st(CO2), Q0st(C2H4), and ETR, with CO2 selectivity constrained by C2H4 binding strength.
- AJOTEY is identified as the optimal MOF candidate with a TSN of 6.43 mol/kg.

## Abstract

The escalation of thermal runaway in lithium-ion batteries presents severe safety hazards that necessitate advanced monitoring protocols to ensure early warning of potential failures. Carbon dioxide (CO2) is released during preliminary decomposition well before catastrophic failure occurs, thereby providing a strategic advantage for early-stage warning. Consequently, identifying materials with high-selective CO2 recognition is an essential prerequisite for developing reliable sensing platforms. This study integrates Grand Canonical Monte Carlo simulations with Random Forest (RF) models to systematically screen 1470 MOFs from the CoRE-MOF 2019 database. The screening process evaluates selective CO2 recognition under multicomponent competitive adsorption conditions involving CO2, C2H4, and O2. The performance evaluation is based on working capacity, selectivity, and the trade-off between working capacity and selectivity (TSN). The RF model achieves high predictive accuracy, with tested R2 exceeding 0.92 on the test samples. Shapley Additive Explanations (SHAP) interpretability analysis identifies Q0st(CO2), Q0st(C2H4), WEPA, KH(C2H4), and ETR as key performance drivers. The results indicate that CO2 selectivity is constrained by the binding strength of competing C2H4. Optimal materials tend to have hard Lewis acid centers and polar inorganic clusters to minimize non-specific π-interactions with interfering species. Top-performing MOFs require balanced structural features, concentrating in moderate surface areas (965–1975 m2/g), narrow pore windows (PLD ≈ 4–7 Å, LCD ≈ 5.5–9.6 Å), high void fractions above 0.6, and low densities below 1.3 g/cm3. AJOTEY emerges as the optimal candidate with a TSN of 6.43 mol/kg, combining substantial working capacity (4.57 mol/kg) with strong selectivity (25.52). These results will accelerate the discovery of sensing materials and provide a practical pathway for MOF-based CO2 sensor development to enhance lithium-ion battery safety.