AI Transforming Battery Research: From Atomic to System Scale

Dr. James Luhoo, a senior lecturer at the University of Greenwich and a fellow of the Faraday Institution, opened an evening at the Royal Institution by drawing parallels between historical scientific breakthroughs and the current impact of AI. He noted that 200 years ago, in the same building, Humphrey Davy pioneered the use of electricity to split matter, discovering fundamental elements like calcium, potassium, and sodium. This transformed electricity from a curiosity into a scientific tool. Davy's assistant, Michael Faraday, despite lacking formal education, became one of the most important experimental scientists, lecturing and working in the same institution. Faraday's laws of electrochemistry are still fundamental to modern battery technology.

Today, the Faraday Institution continues this legacy, focusing on energy storage and movement. The elements discovered by Davy, such as lithium and sodium, are now crucial components of batteries.

The Microscopic World of Batteries

While batteries may appear unexciting externally, their internal microscopic world is incredibly complex. Ions shuttle between electrodes thousands of times, causing swelling, cracking, and degradation, which limits battery life. Strange films grow on surfaces, and hydrogen evolution can occur. This intricate interplay of physics, chemistry, and mechanics makes batteries a rich area of scientific study.

The Role of Artificial Intelligence

Artificial intelligence (AI) is becoming indispensable in understanding and improving batteries. It helps unravel individual degradation mechanisms and mechanics, which are otherwise challenging to decipher. AI can piece together information from disparate sources and generate new material classes by combining redox couples across the periodic table, accelerating scientific discovery.

The evening featured three presentations exploring AI's application at different scales of battery research:

1. AI at the Atomic Scale: Designing New Materials

Professor Aron Walsh, Chair of Computational Materials Design at Imperial College and Chief Scientific Officer of Casper AI, discussed using AI as a "digital alchemist" to design new materials.

For millennia, humanity has sought to understand the building blocks of materials. Ancient Greeks believed materials were combinations of classical elements, with Plato even assigning geometric features to them. In the 17th century, Robert Boyle proposed the existence of indivisible "atoms." John Dalton later arranged elements by relative weight, and Mendeleev developed the periodic table, which allowed for predictive chemistry.

Today, with centuries of modern chemistry, millions of published papers, chemical compositions, and characterized structures, the sheer volume of data is beyond human processing capacity. This is where machine learning excels. AI can encode this vast chemical information into numerical vectors, allowing computers to identify patterns and similarities. Different AI models, trained on diverse data, consistently extract a common chemical picture, akin to Mendeleev's observations.

The potential material space is enormous. Considering combinations of elements and their spatial arrangements, the number of possible materials far exceeds the estimated number of atoms in the universe (10^100 vs. 10^80). This vastness makes traditional exploration impossible. AI, however, can learn distributions and efficiently target promising regions.

The Ideal Battery of the Future: * Sustainable components: Low-cost, environmentally friendly elements. * Increased power: Enabling applications like electric planes. * Solid-state batteries: Eliminating flammable organic electrolytes. * Accelerated development: Leveraging current knowledge and tools to shorten commercialization times.

Generative AI for Materials: Generative AI, like that used to create images from text prompts, can be applied to materials design. By providing context from vast datasets of known materials, compositions, structures, and performance data, AI can generate new materials that meet specific criteria (e.g., a cobalt-free cathode for a sodium-ion battery with high voltage and capacity).

Walsh's group developed "Chameleon," a text-to-material model, as a proof of concept. More recently, "co-scientists" like "Crystallize" (developed by Ryan Anduma) are emerging. These AI agents can review literature, propose hypotheses, perform experiments (through simulations or robotic labs), and evaluate results, iterating to refine solutions.

Walsh co-founded Cusp AI, an AI for materials startup, to scale these capabilities. Their "Mira" co-scientist acts as a materials intelligence engine. While current AI models are excellent at analyzing existing data and trends, they sometimes lack the "confidence" to explore entirely new chemical spaces beyond known materials. This highlights the ongoing need for human expertise in building tools and integrating AI with experimental labs.

AI has transformed materials theory and simulation, automating routine tasks, enabling exploration of vast material spaces, and orchestrating complex research campaigns in hours or days.

2. AI at the System Scale: Understanding Battery Performance and Degradation

Dr. Mona Faraji Nery, Associate Professor at the University of Warwick and an Alan Turing Fellow, discussed how AI acts as a "forensic investigator" to understand battery performance and degradation.

Batteries degrade from their first use due to mechanical stress, particle cracking, lithium plating, and dendrite formation, leading to reduced capacity and power over time. This degradation is influenced by numerous factors: * User habits: Aggressive driving, frequent full charges/discharges. * Chemistry: Different chemistries degrade differently. * Cell format: Cylindrical, pouch, or coin cells exhibit varying degradation. * Manufacturing process: Over 114 steps and 600 control factors (temperature, mixing speeds, etc.) impact final battery quality. * Environmental factors: Temperature fluctuations, humidity, dust, saline water, and idle time.

Given the multitude of influencing factors, predicting a battery's remaining useful life (RUL) is a "million-dollar question." The demand for batteries, particularly in electric vehicles, is immense. For example, the projected annual demand for EV batteries in the UK by 2035, if stacked as cylindrical cells, would stretch from the Earth to the Moon and back.

No human can link all these factors to battery degradation. This is where AI excels. Batteries generate vast amounts of data (temperature, current, voltage, images, microstructures) at different scales. AI technologies, especially machine learning and deep learning, can identify underlying patterns between these factors and battery degradation.

How AI Works: 1. Feature Extraction: AI processes raw data, extracting critical information. 2. Model Training: Models are trained to link these features to RUL, using techniques like regression, time series modeling, and attention mechanisms (transformers). 3. Physics-Informed AI: Combining data-driven AI with established physics (differential equations) provides a more robust understanding of degradation.

The goal is to predict battery lifespan quickly (e.g., in a month or two) rather than waiting years, which is crucial given the rapid pace of material and manufacturing advancements. AI models need to be transferable across different conditions, users, and regions, while ensuring privacy, security, and accuracy. Embedding complex AI algorithms into hardware for efficient use in vehicles or mobile phones is an ongoing challenge.

AI helps both understand how batteries degrade and improve manufacturing processes to create better, next-generation batteries. This is currently "narrow AI," focused on specific problems. The future involves "general AI" with agents that can autonomously gather information and "super AI" with advanced reasoning capabilities. The success of AI in this field depends on data, computational power, and model complexity, representing the intersection of hardware, software, and science.

3. AI at the Microstructure Scale: Bridging Atoms and Devices

Dr. Sam Cooper, Associate Professor of AI for Materials Design at Imperial College London and Chief Scientific Officer of Polaron, focused on the critical but less understood realm of microstructure.

Microstructure refers to the internal structure of materials at the micro-scale (10 nanometers to 100 microns), including pores, particles, phases, and interfaces. Unlike atomic structures, microstructures emerge indirectly from manufacturing processes, making them difficult to predict or directly specify.

Why Microstructure Matters: In a lithium-ion battery, lithium ions must travel through tortuous pore networks between anode and cathode particles. Improving the shape of this network can significantly enhance cell performance, leading to faster charging and longer driving ranges.

Manufacturing Process: Battery electrodes are typically made through four steps: 1. Mixing: Active materials, binders, additives, and solvents are vigorously mixed. 2. Coating: The resulting "ink" is finely coated onto a metallic foil. 3. Drying: Solvents evaporate in an oven. 4. Calendering: The coated foil is crushed by rollers to increase density.

These processes involve complex physics (solids, liquids, gases, fluid flow, evaporation, phase change, fracture, deformation), resulting in intricate microstructures, as seen in scanning electron microscopy (SEM) images. Traditional physics-based modeling struggles to capture this complexity, often relying on idealized representations.

AI's Contribution to Microstructure: Cooper outlined four ways AI is improving microstructure understanding:

1. Characterization: AI, particularly computer vision models like Meta's DinoV2, can rapidly and accurately segment SEM images, identifying different materials (active material, pore, carbon binder domain) at a pixel level. These models, trained on vast amounts of natural images, develop an internal representation of the world that can be applied to microstructural data.

2. Generation: Many critical battery properties are 3D. Generative AI can create 3D microstructures from a single 2D image, significantly accelerating characterization workflows and allowing for the extraction of complex 3D metrics (particle size distribution, pore network tortuosity, electronic transport).

3. Design: By training generative models on images from materials manufactured with different parameters, AI can explore the space of all possible manufacturable electrodes. This enables optimization loops to identify manufacturing parameters that yield desired microstructures and battery performance. This led to the spinout company Polaron. Microstructure is also crucial in pharmaceuticals, alloys, and catalysts.

4. Validation: Robotic labs are increasingly used for repetitive tasks, ensuring consistency and collecting comprehensive data (weight, photographs) at every stage. Agentic AI models, trained on scientific literature, can design experimental workflows for these robotic labs, reacting in real-time to emerging data and making intelligent decisions about subsequent steps.

Microstructure is the "missing layer" that connects atomic possibilities with device necessities. AI is making this layer visible, designable, and controllable, paving the way for next-generation batteries.