An Alternative to LLM-based AI for Engineering Applications

Almost all of the AI that is currently deployed makes use of large language models (LLM). As the name suggests, this type of algorithm is based on processing language. And while LLMs may excel in tasks such as creating text, audio, images and video, its usefulness diminishes quickly when making calculations based on numerical data, such as mathematical modelling, predictive modelling, and statistical analysis. According to Dr. Ang Xiao, technical lead of AI for materials science at SandboxAQ, large quantitive models (LQMs) are a much better tool for that type of application. Xiao has just presented a paper at the , which won the award for AI Battery Startup of the Year. The paper contained two separate case studies demonstrating how LQM-based AI had the capability to advance battery technology

He explained, “Today's AI focuses on the use of LLMs to manipulate text, images and videos. This is helpful, but at SandboxAQ, we believe that future AI will rely on its capabilities to solve more challenging problems, for example, in chemistry, physics and biology. We will have to task the AI to discover new drugs, or find new battery chemistries to make a sustainable future. So, we believe this new type of AI model has the potential to unlock trillions of dollars of market opportunities across many industries - in aerospace, automotive, energy, chemical materials, biopharma and many others.”

SandboxAQ is a pioneer in the development of LQMs that are grounded in physics and built to simulate real-world systems. The company was spun out of Alphabet in 2022, and since then has raised almost $1 billion in investment. Its business strategy is to build and train AI models that will help their customers to solve their own problems rather than attempting to make new discoveries themselves. Customers approach them with technical problems, which SandboxAQ solves using its proprietary AI and simulation technologies. To reassure the customer, SandboxAQ does not need access to proprietary customer data, rather, it generate synthetic datasets through advanced simulations.

The first of the two case studies that Xiao presented at the Battery Show concerned the prediction of battery lifetime. Batteries degrade, or fade, over time, and each application has a figure when they are considered spent. For EVs that number is usually between 70% and 80% of their original capacity. Other applications, such as renewable energy, have lower thresholds. Each different battery chemistry degrades at different levels, and that degradation figure is hugely valuable information for OEMs, cell makers and users, who need to know the expected lifetime of the battery under its unique usage conditions, without waiting several years to let the battery cycle to the end of life. The fade figure of the battery can be used to calculate the lifetime, but the calculations are complex.

Xiao says, “The traditional model of predicting battery fade measures a capacity versus voltage curve over two battery cycles, and uses the delta to predict the capacity. Toyota and MIT published a paper several years ago in Nature for their own model. It works well for LFP batteries, but it can’t be used to predict other cell chemistries, and it can’t predict remaining useful life reliably. The model is not broad or not general enough. The method used by SandboxAQ requires additional details, such as the electrochemical impedance, resistance, and dQ/dV analysis on a particular cycle, and we transfer them as features to the trained AI, which then predicts the parameters and their weighting to enter into physical equations that allow us to predict the end of life for all different cell chemistries across different manufacturers.”

The second case study presented by Xiao at the Battery Show concerned new material discovery, specifically targeting new cathodes and solid-state electrolytes. The SandboxAQ models use machine learning ab initio simulations to understand atomic-level interactions within potential battery materials. They use the neural networks to predict fundamental properties (such as voltage, structural stability, and ion mobility), enabling them to sift through huge databases and identify materials that meet advanced design requirements, for example being cobalt-free, to mitigate supply-chain concerns. Using this technique, the company has already successfully identified and validated three new promising cathode materials.

These two case studies are only a taste of the potential benefits that LQMs can bring to the battery market. There are also many other projects ongoing with a wide range of customers. The potential benefits of AI are now being realized.

https://www.sandboxaq.com/