Source-linked AI summary
Machine Learning the Voltage of Electrode Materials in Metal-ion Batteries
Rajendra P. Joshi, Jesse Eickholt, Liling Li, Marco Fornari, Veronica Barone, Juan E. Peralta
TL;DR
The study addresses challenges in scaling lithium-ion batteries, including low energy density and uncertain lithium availability, by applying machine learning to electrode-voltage prediction. Its models achieve mean absolute errors around 0.39–0.43 V across reported test sets, while limited data constrains performance in some ranges.
Problem
Scaling lithium-ion batteries remains challenging because of relatively low energy density and the scarcity of lithium as a raw material.
Method
The paper machine-learns electrode-material voltages using deep neural networks, support vector regression, and kernel ridge regression with elemental-constituent properties.
Results
Mean absolute errors were 0.43 V on the H-set, 0.42±0.13 V on the T-set, and 0.39 V on another H-set evaluation.
Takeaways & Limitations
Machine learning provides a demonstrated approach for estimating electrode voltages in metal-ion batteries.
Takeaways & Limitations
Limited data may be insufficient for training the model in some voltage ranges, and comparatively poorer performance was associated with a large intercept.
Abstract
from arXiv · showhide
Machine learning (ML) techniques have rapidly found applications in many domains of materials chemistry and physics where large data sets are available. Aiming to accelerate the discovery of materials for battery applications, in this work, we develop a tool (http://se.cmich.edu/batteries) based on ML models to predict voltages of electrode materials for metal-ion batteries. To this end, we use deep neural network, support vector machine, and kernel ridge regression as ML algorithms in combination with data taken from the Materials Project Database, as well as feature vectors from properties of chemical compounds and elemental properties of their constituents. We show that our ML models have predictive capabilities for different reference test sets and, as an example, we utilize them to generate a voltage profile diagram and compare it to density functional theory calculations. In addition, using our models, we propose nearly 5,000 candidate electrode materials for Na- and K-ion batteries. We also make available a web-accessible tool that, within a minute, can be used to estimate the voltage of any bulk electrode material for a number of metal-ions. These results show that ML is a promising alternative for computationally demanding calculations as a first screening tool of novel materials for battery applications.
Introduction
The work addresses the difficulty of screening a large space of candidate electrode materials for emerging metal-ion batteries. It develops machine-learning models and a web-accessible tool to predict electrode voltages efficiently.
- Emerging Na-, K-, Mg-, Ca-, and Al-ion batteries need suitable electrode and electrolyte materials, while many possible electrode compounds remain unexplored.
- Machine learning is proposed to screen the large chemical and structural space of possible electrode materials more quickly and with limited computational use.
- The study develops a tool that predicts electrode voltages for metal-ion batteries using Materials Project data.
- The models combine compound and elemental-property feature vectors with deep neural networks, support vector machines, and kernel ridge regression.
- The work compares machine-learning voltage profiles with density functional theory and provides a web interface for estimating voltages within minutes.
- The authors conclude that machine learning can efficiently predict electrode-material voltages as an exploratory screening tool.
Data and Feature Vectors
The dataset contains intercalation-electrode voltage instances from the Materials Project and represents each compound using structural, chemical, and elemental features. Feature normalization and principal component analysis reduce redundancy before model training and evaluation.
- The dataset contains 4,250 voltage instances for 3,580 intercalation-based electrode materials and includes DFT-predicted voltages for several metal-ion batteries.
- After removing inconsistencies and repetitions, 3,977 data instances remained, with average voltages defined across intercalation ranges.
- Each material is represented by its working ion, active-ion concentration, crystal lattice, space group, and elemental properties of its constituents.
- The feature construction produces 237 features, whose inputs are normalized between -1 and 1 to avoid magnitude-driven feature preference.
- Principal component analysis reduces feature-space dimensionality, with 80 principal components retained for the machine-learning models.
Results and Discussion
Across cross-validation and holdout tests, the ML models predict electrode voltages with moderate error, while performance worsens on Na-set data and low-voltage materials. The models also reproduce DFT voltage trends, support screening Na- and K-ion electrodes, and underpin a public voltage-prediction tool.
- Model performance: The models achieved mean MAEs of 0.43 V for DNN on the T-set and H-set, while KRR achieved 0.46±0.14 V on the T-set.The DNN T-set result had a standard deviation of ±0.03 V.
- Model performance: Models performed slightly worse at lower voltages, which the authors attribute to limited data, and suggest could improve with more low-voltage training points.This limitation affects the data coverage used for training rather than establishing a general failure of the models.
- Model performance: R2 values of 0.84, 0.86, and 0.86 were obtained for DNN, SVR, and KRR, respectively, on the H-set.The H-set MAEs were similar across algorithms, including 0.39 V for KRR and 0.40 V for SVR.
- Transferability: Na-set performance was poorer than T- and H-set performance, with DNN error about 0.8 V larger, partly because Na training data were scarce and the test set contained only 32 materials.The Na-set retained a linear predicted-versus-target relationship despite its larger error.
- Applications: The models identified many potential Na- and K-ion electrodes and were deployed as a web tool for predicting average voltages of several metal-ion systems.The tool is publicly accessible and predicts voltages from minimal basic information for novel electrodes.
Conclusions
The study uses ML models trained on Materials Project DFT data to predict electrode voltages, reproducing DFT trends and enabling rapid screening of battery materials. It proposes nearly 5,000 Na- and K-ion electrode candidates and provides a web tool for voltage prediction, while noting data and performance limitations.
- Conclusions: The models machine-learn electrode voltage from DFT data extracted from the Materials Project Database and evaluate performance using mean absolute error across training, holdout, and literature data.
- Conclusions: ML models reproduce DFT voltage trends and can guide rapid screening and experiments for novel battery materials.The methodology is faster than DFT and can support screening before synthesis.
- Conclusions: Nearly 5,000 electrode materials are proposed as candidates for Na- and K-ion batteries.
- Conclusions: Routine voltage prediction may require improved models, more data, alternative algorithms, and broader representation of intercalation reactions.Robust training also requires DFT records spanning all voltage ranges, including unsuitable electrode materials as negative instances.
- Conclusions: Including negative data instances in databases is necessary for proper training of data-intensive models.
- Conclusions: A web-accessible tool predicts the voltage of any electrode material within a minute.
Supporting Information Available
The supporting information summarizes visual and tabular comparisons of ML voltage predictions with DFT values across test data. It includes model-performance plots, parameter tuning information, and Na- and K-based electrode results.
- Supporting Information Available: Scatter plots show the performance of several ML models on holdout test sets and the Na test set.
- Supporting Information Available: A color map presents parameter tuning for Li-only data.
- Supporting Information Available: The supporting information compares predicted voltages from DNN, SVM, and KRR with DFT values for Na-based electrodes.
- Supporting Information Available: It reports ML-predicted and DFT voltages for Na- and K-based electrodes in a table.