Google, Carnegie Mellon Use AI for Battery Breakthroughs

Electrification depends on the advancement of battery technology. To meet demand, researchers are using artificial intelligence (AI) and machine learning to discover more effective and efficient materials. 

 

AI research can improve EV batteries. Image used courtesy of Pixabay and Wikimedia Commons 

 

In separate studies, teams from Google Deep Mind and Carnegie Mellon used deep learning to make breakthroughs with the potential to advance electric vehicle batteries, lithium-ion technologies, and battery energy storage systems.

 

Google DeepMind Discovers Battery Applications

Recently, Google DeepMind researchers employed deep learning to discover millions of new materials with possible applications in solid electrolytes, solid-state batteries, conductors, and electric vehicle (EV) batteries.

The team found over 2.2 million crystals and 380,000 stable materials. These materials are promising candidates for experimental synthesis. This advancement is a major leap, representing an order-of-magnitude increase compared to previous discoveries, bringing the total to 421,000 stable crystals. Within this, the group found 528 potential lithium-ion conductors, some of which could be used to improve the performance of lithium-ion batteries. These breakthroughs are akin to adding nearly 800 years' worth of knowledge to the field of material science.

 

The structural pipeline of GNoME. Image used courtesy of Google DeepMind

 

DeepMind's breakthrough used its AI tool, Graph Networks for Materials Exploration (GNoME). GNoME utilizes a state-of-the-art graph neural network model to discover low-energy (stable) materials. The tool is originally trained with data on crystal structures and their stability. It is available through the Materials Project and operates through two pipelines: a structural pipeline that creates candidates with structures similar to known crystals and a compositional pipeline following a more randomized approach based on chemical formulas. The performance of GNoME is continually enhanced through a process known as “active learning,” in which the model's predictions are tested and refined using established computational techniques.

GNoME employs an innovative approach combining model-based filtration and density functional theory calculations. The result is that GNoME's deep learning models accurately predict crystal energies to within 11 meV/atom and have achieved a stable prediction precision of over 80% for structures and 33% per 100 trials for compositions, starting from less than 6% and 3%, respectively, in earlier stages​​. With this knowledge, GNoME can discover stable crystals with potential value in power electronics applications like EV batteries, superconductors, and semiconductors.

In addition to making these predictions, Google DeepMind has made GNoME’s findings available to the research community. These materials are being contributed to the Materials Project and added to its online database for further processing and analysis. 

 

CMU Uses AI for Autonomous Chemical Research

In a similar vein, researchers at Carnegie Mellon University designed an AI system to autonomously plan, execute, and optimize scientific experiments. The system, dubbed Coscientist, is developed around OpenAI’s GPT-4 large language model and has been tuned for the automation of scientific discovery, particularly within chemical research. This research can advance energy storage technology.

Coscientist interacts with multiple modules for problem-solving, including web and documentation search, code execution, and direct experimentation. Its architecture consists of a primary module called “Planner,” a GPT-4 chat completion instance that plans based on user inputs by invoking various commands like GOOGLE, PYTHON, DOCUMENTATION, and EXPERIMENT. These commands allow it to gather information, perform computations, and conduct experiments. For example, the GOOGLE command enables it to perform web searches, and the DOCUMENTATION command aids in retrieving and summarizing necessary documentation for experiment execution.

 

An overview of documentation search with Coscientist. Image used courtesy of Boiko et al.

 

The team tested the system across six diverse tasks, demonstrating its ability to plan chemical syntheses, search through hardware documentation, execute high-level commands in a cloud laboratory, control liquid handling instruments, tackle complex scientific tasks integrating multiple hardware modules, and solve optimization problems based on previously collected experimental data. It has shown considerable promise in performing these tasks with efficiency and accuracy, highlighting the potential of AI-driven autonomous systems in scientific research methodologies.

Coscientist's capabilities extend to understanding and programming in the ECL Symbolic Lab Language (SLL), enabling it to interface with complex laboratory instruments and perform high-level chemical research tasks. This method could have major implications in fields such as battery development, which rely on the precise manipulation and analysis of materials at the molecular level. By leveraging SLL, Coscientist could automate intricate experiments, optimize testing protocols, and accelerate the discovery of chemical compounds and energy storage solutions.

 

Deep Learning for Science

The advancements from Google Deep Mind and Carnegie Mellon University signify a transformative phase in material science research driven by integrating deep learning technologies. These advances could have major implications in the battery industry, which relies on material advancements and unique chemistries to achieve greater capacity, power output, safety, and reliability.