Chemistry-Informed Machine Learning for Polymer Electrolyte Discovery

Gabriel Bradford, Jeffrey Lopez, Jurğis Ruža, Michael A. Stolberg, Richard Osterude, Jeremiah A. Johnson, Rafael Gómez‐Bombarelli, Yang Shao‐HornACS 2023 · DOI

TL;DR

Researchers have developed "ChemArr," a chemistry-informed machine learning model that accurately predicts ionic conductivity in solid polymer electrolytes (SPEs). By encoding the Arrhenius equation, ChemArr significantly outperforms models that don't account for temperature dependence, accelerating the discovery of promising new materials for safer, higher-energy-density batteries.

What Problem Does This Paper Address?

Lithium-ion batteries are crucial for energy storage, but current liquid electrolytes pose safety risks (flammability) and limit energy density. Solid polymer electrolytes (SPEs) offer a safer alternative but historically suffer from low ionic conductivity, hindering their widespread adoption. Discovering new SPEs with high conductivity is a slow and costly process involving extensive experimental trial-and-error. Existing machine learning approaches often struggle to generalize due to limited data or a lack of incorporation of fundamental physical principles governing ionic conductivity.

Methods

The core of this work is the ChemArr model, a novel machine learning architecture. It builds upon a message passing neural network (MPNN), similar to the Chemprop model, to learn representations of polymer and salt structures directly from their molecular graphs. What makes ChemArr unique is its chemistry-informed readout layer. Instead of directly predicting ionic conductivity, the MPNN's penultimate layer outputs parameters for the Arrhenius equation ($ln(A)$ and $E_a$). These parameters, along with the input temperature, are then fed into a fixed-weight layer that calculates the ionic conductivity ($\sigma$) using the Arrhenius equation: $ln(\sigma) = ln(A) - E_a / (RT)$. This explicit incorporation of the Arrhenius equation, which describes temperature-activated processes, is the key "chemistry-informed" aspect. The model was trained using a mean squared error loss function.

Datasets

The primary dataset used for training and evaluation was compiled from over 200 experimental publications on SPE ionic conductivity. This dataset comprises 12,383 data points covering 10,272 unique SPE formulations. The authors also incorporated data from two previously published datasets, adding an additional 6,250 data points. For benchmarking, the authors created test and validation sets that excluded specific polymers to assess the model's ability to generalize to unseen polymer chemistries. A screening set of over 20,000 hypothetical SPE formulations was generated for discovery purposes.

Key Insights and Ideas

  • The Power of Physics-Informed ML: Explicitly encoding the Arrhenius equation into the ML model significantly improved prediction accuracy and generalizability. This highlights the benefit of integrating known physical laws into data-driven models, especially when experimental data is limited.
  • Beyond Molecular Fingerprints: Using molecular graphs directly as input to the MPNN allows the model to learn optimal representations without relying on pre-defined, potentially information-losing, molecular fingerprints.
  • Anion Chemistry Matters, But How Depends on the Polymer: The study revealed that anion properties (molecular weight, interaction energy) strongly influence ionic conductivity. Crucially, the nature of this influence is reversed between poly(ethylene oxide) (PEO) and poly(trimethylene carbonate) (PTMC), suggesting that polymer-salt interactions are complex and polymer-specific.
  • Identifying Promising New Materials: The model successfully screened a vast chemical space, identifying promising polymer structures (e.g., siloxane-based and phosphazene-based polymers) that warrant further experimental investigation.

Key Results

  • Improved Accuracy: ChemArr achieved a mean absolute error (MAE) of 1.00 ± 0.030 log(S/cm), outperforming XGBoost (1.09 ± 0.027) and Chemprop (1.08 ± 0.012) in cross-validation.
  • Enhanced Ranking: ChemArr showed a 32% improvement in Spearman R (0.59 ± 0.022) compared to Chemprop (0.45 ± 0.006), indicating a better ability to rank the ionic conductivity of different SPE formulations.
  • Low Conductivity Prediction: ChemArr's predictions were particularly accurate for low ionic conductivity values (below 10⁻⁸ S/cm), where other models struggled.
  • Generalizability: The model predicted near experimental accuracy for 49% of all polymer species in the test set and was within 1 order of magnitude of experimental data for 81% of SPEs.
  • Discovery Potential: Over 20,000 candidate SPE formulations were screened, identifying several promising polymer structures for further research.
  • Anion Trends: For PEO, anions with moderate molecular weight (around 300 g/mol) and low interaction strength were favored. For PTMC, the trend reversed, favoring lower molecular weight and higher interaction strength anions.

Limitations and Future Work

  • Non-Arrhenius Behavior: ChemArr struggled with SPEs exhibiting non-Arrhenius behavior (e.g., due to phase transitions or decomposition), leading to higher errors at low temperatures. While attempts were made to use the Vogel-Fulcher-Tammann (VFT) equation, it did not fully resolve this limitation.
  • Data Sparsity: Predictions for polymers with limited representation in the training data (e.g., those not coordinating lithium with oxygen) showed increased error.
  • Experimental Data Reliability: In some cases, high prediction errors were attributed to potentially unreliable experimental data (e.g., due to polymer decomposition).
  • Future Work: The authors suggest extending ChemArr to other processes with Arrhenius-type dependencies and exploring the use of physics-informed ML in other domains where data is scarce. Further experimental and simulation work is needed to confirm the observed anion trends and understand the underlying mechanistic reasons for the polymer-specific behavior. The development of models that can capture non-Arrhenius behavior is also a key area for future research.

Summary updated

Abstract

Solid polymer electrolytes (SPEs) have the potential to improve lithium-ion batteries by enhancing safety and enabling higher energy densities. However, SPEs suffer from significantly lower ionic conductivity than liquid and solid ceramic electrolytes, limiting their adoption in functional batteries. To facilitate more rapid discovery of high ionic conductivity SPEs, we developed a chemistry-informed machine learning model that accurately predicts ionic conductivity of SPEs. The model was trained on SPE ionic conductivity data from hundreds of experimental publications. Our chemistry-informed model encodes the Arrhenius equation, which describes temperature activated processes, into the readout layer of a state-of-the-art message passing neural network and has significantly improved accuracy over models that do not encode temperature dependence. Chemically informed readout layers are compatible with deep learning for other property prediction tasks and are especially useful where limited training data are available. Using the trained model, ionic conductivity values were predicted for several thousand candidate SPE formulations, allowing us to identify promising candidate SPEs. We also generated predictions for several different anions in poly(ethylene oxide) and poly(trimethylene carbonate), demonstrating the utility of our model in identifying descriptors for SPE ionic conductivity.
BibTeX
@inproceedings{bradford2023chemistry,
  author = {Gabriel Bradford and Jeffrey Lopez and Jurğis Ruža and Michael A. Stolberg and Richard Osterude and Jeremiah A. Johnson and Rafael Gómez‐Bombarelli and Yang Shao‐Horn},
  title = {{Chemistry-Informed Machine Learning for Polymer Electrolyte Discovery}},
  booktitle = {ACS},
  year = {2023},
  doi = {10.1021/acscentsci.2c01123},
}
Figure