Editor's note: Lithium batteries have a long lifespan. We often hear "drive out, push the cart back", "charge for two hours, standby for two minutes", and also hear "the temperature drops sharply, please pay attention to keeping the battery warm"... As the Products powered by lithium-ion batteries, such as mobile phones, computers, new energy vehicles, etc., have gradually become necessities in people's lives, and there are more and more concerns about battery life. Battery performance prediction has also become one of the important topics in industrial artificial intelligence research.
In order to better analyze battery performance and predict battery service life, Microsoft Research Asia developed and open sourced the one-stop machine learning tool BatteryML, hoping to gather more professional forces to jointly promote research in the battery field.
In recent years, lithium-ion batteries have become the cornerstone of energy storage solutions due to their high energy density, long cycle life and relatively low self-discharge, and are also widely used in various commercial scenarios, including new energy vehicles, consumer electronics and energy storage facilities, etc. Although lithium batteries bring many advantages, they still face challenges such as capacity fading and performance optimization.
In the process of continuous recycling, the inherent electrochemical characteristics of lithium batteries inevitably lead to a decline in their performance, which is specifically manifested as a decrease in charge and discharge capacity. This uncontrolled performance degradation will have a great impact on downstream business scenarios, such as causing "mileage anxiety" for new energy vehicle users, affecting the power supply stability of energy storage systems, and so on. Moreover, excessive capacity decay of lithium batteries will also bring great challenges to sustainable development, including increased equipment maintenance costs, consumption of scarce resources, aggravated environmental pollution, and impact on industrial economic benefits. Therefore, effectively analyzing and predicting the performance degradation of lithium batteries, and thus providing guidance for early prevention and intervention, has become a very important industrial artificial intelligence research topic.
Overcome modeling challenges to analyze and predict battery performance
The performance degradation of lithium batteries is a complex electrochemical process, which involves the growth of solid electrolyte membrane, lithium precipitation, loss of active materials, etc. This process will be affected by multiple factors such as electrode materials, ambient temperature, charge and discharge conditions and rates, making it difficult for physical models based on limited electrochemical laws to effectively model the battery performance degradation process under actual conditions. However, today's machine learning methods can automatically summarize complex patterns from data, and have become an important tool for modeling battery performance degradation in recent years, attracting widespread attention from academia and industry.
However, researching and applying machine learning models in battery performance modeling scenarios is also extremely challenging. On the one hand, domain experts in the battery industry, although they have a deep understanding of the mechanisms and principles behind battery degradation, are not good at building effective machine learning models. Because building a machine learning model often requires special data processing, effective feature construction, precise model construction and tuning and other preparations. On the other hand, for data scientists in the computer industry, the battery field has characteristics such as severe data heterogeneity, high domain knowledge requirements, and diverse task definitions, which greatly hinders their research and application of the most advanced machine learning methods.
To address these challenges, researchers at Microsoft Research Asia developed the BatteryML tool for battery performance analysis and prediction. At the same time, Microsoft Research Asia also hopes that BatteryML can not only become a one-stop solution for experts in the battery industry, but also serve as an efficient development platform for computer scientists to study battery performance prediction.
For detailed documentation, examples and source code of BatteryML, please visit the BatteryML GitHub link: https://github.com/microsoft/BatteryML
Building a community-driven open source platform
As a one-stop solution, BatteryML simplifies the research and experimental process of battery data, covering various classic models in the field of lithium battery research.
BatteryML architecture diagram
BatteryML has six main features:
- Open source and community-driven: BatteryML aims to create a community-driven open source platform to bring together experts in the battery field and data scientists to jointly promote progress in the field of battery performance modeling.
- Unified data representation: BatteryML builds a unified set of data representations to deal with battery data heterogeneity and provides comprehensive processing scripts to summarize almost all currently public battery data sets.
- Preprocessing and feature engineering: BatteryML provides a set of standard data preprocessing processes by default, and has built-in basic feature engineering to meet the high requirements of domain knowledge.
- Rich models: BatteryML covers existing classic models and benchmarks in the field of battery performance prediction.
- Clear task system: BatteryML includes the battery performance prediction tasks that are currently the most concerned, which are both aligned with the tasks in classic research papers and include some variant tasks that are more concerned in the industry.
- Extensible and customizable: BatteryML reserves a comprehensive interface so that researchers and developers can efficiently customize unique data sets, develop new data processing and feature engineering, and study more advanced machine learning according to their own needs. Models etc.
In the future, Microsoft Research Asia's research progress on battery performance prediction will be open sourced on the BatteryML platform, including the previously proposed "multi-faceted deep contrastive regression" method and the application of large models in the battery field.
Through the open source BatteryML solution, Microsoft Research Asia looks forward to accelerating the integration of industry, academia and research in the field of battery performance prediction, and hopes that more partners who are concerned about battery performance modeling will join, contribute new functions, new module codes, and share new open source data with the community. , jointly promote research progress in the battery field.