Battery manufacturers commit millions of dollars to materials, suppliers and production processes long before they know how the cells will age. Establishing lifetime performance can take years, and waiting for that evidence is rarely practical. Yet moving ahead without it can lead to costly redesigns or production changes if performance falls short.
That uncertainty slows innovation and makes promising technologies harder to bring to market. A new material may lower production costs or ease a supply constraint, but a shorter battery life can easily outweigh those benefits. Manufacturers need a way to assess that risk while they still have room to change course.
Artificial intelligence helps teams make better use of the evidence available early in development. By analyzing the first cycles of testing, predictive models can identify signs of degradation and offer insight into how a cell ages while longer tests continue. That may be enough to justify a pilot run or flag a concern worth investigating before a supplier contract is signed.
Putting those forecasts to work requires understanding what the early measurements reveal and where uncertainty remains. With that perspective, teams can focus limited testing resources on the most promising options and make better-informed decisions before committing to production.
Here’s what that looks like in practice.
Look Beyond Early Capacity Loss
A cell can retain most of its original capacity even as changes inside it begin to affect its long-term performance. Reactions at the electrode surfaces and deterioration within the electrodes can make it harder for charge to move through the cell. Those changes can show up as increased internal resistance before a conventional capacity test reveals much deterioration.
Engineers can investigate this behavior using direct current internal resistance (DCIR) measurements. They apply a controlled current pulse and measure the resulting voltage change, repeating the test at different states of charge. Tracking these responses over successive cycles adds information that capacity measurements alone can miss. Combined with voltage-capacity data, these early signals can help predictive models distinguish how different cell designs and electrode materials will age.
In research I coauthored with researchers at NOVONIX, we examined 57 commercial lithium-ion cells from three manufacturers. The study reported cycle-life outcomes ranging from roughly 700 to 2,400 cycles, depending on the capacity threshold used. We built machine-learning models using data from the first 102 cycles, including DCIR measurements that many cell makers already collect. Models incorporating these measurements predicted cycle life more accurately across cell types than those relying only on conventional voltage-capacity features.
These findings are encouraging, although they reflect our study conditions. All the cells used nickel-based cathodes and graphite-based anodes, so we cannot assume similar predictive performance for other chemistries without further validation.
For manufacturers comparing materials or cell designs, early predictions can help determine which candidates warrant more extensive testing while lifetime data is still being collected. Engineering teams can use those forecasts to prioritize promising options and investigate signs of degradation before committing more resources. The value lies in making the next round of testing more focused and informative.
Match the Prediction to the Decision
Our work with NOVONIX showed that early resistance measurements improved cycle-life predictions across the cell types we studied. We have also secured two U.S. patents covering battery cycle-life prediction and early identification of cells likely to experience catastrophic capacity fade. For manufacturers, the next question is how to use that insight when choosing materials and suppliers.
Suppose we’re evaluating a material from an alternative supplier. The first cells meet capacity specifications, and our procurement team sees a chance to lower costs or reduce dependence on a single source. Early measurements can help us assess whether those cells are likely to retain their performance over time, provided the model is applicable to the cells being evaluated.
Newer approaches, such as Large Quantitative Models (LQMs) that combine machine learning with physics-based simulation, can help researchers evaluate material properties before committing to extensive laboratory testing. Both cell-life prediction and materials modeling can help teams decide which candidates warrant further testing and what that testing needs to establish.
To compare the new material fairly, test cells made with it under the same conditions as cells made with the existing supply. Keep each cell’s results tied to its material lot and production settings so differences can be investigated. Ask the modeling team how well its training data represents the cells being evaluated. If the new material or test conditions differ substantially from those represented in the training data, a favorable forecast needs more experimental support.
Before test results arrive, procurement, engineering and quality teams should agree on what evidence would justify a larger pilot and what concerns would require more testing (or rule out the material entirely). Put those criteria in writing, along with who will approve the next step. This helps teams avoid approving a supplier based on a promising prediction alone, even when purchasing deadlines are approaching.
As longer-term results arrive, check early predictions against the performance of cells made using the intended production process. Reassess the model when materials or processes change.
Safety requires separate testing. A cycle-life forecast cannot establish how a cell will respond to overcharging, crushing or overheating, or rule out manufacturing defects.
Give Shelf Life a Place in Procurement
When evaluating a new supplier, it’s easy to focus on how cells perform through repeated charging and discharging. But a battery may spend months in storage before it ever reaches a customer, and it continues to age during that time. Qualification testing should account for how storage affects the performance customers can expect when they put the battery to use.
Calendar aging is the deterioration that occurs over time, even when a cell is not being cycled, and our work with the U.S. Armyshows why it matters. The program evaluated batteries against shelf-life requirements ranging from two to 20 years. To understand how storage affects their performance, we tested several chemistries and cell formats at different temperatures and for varying storage durations, alternating periods of storage with cycling. This reflects how batteries move back and forth between active service and storage over their lifetime.
Commercial applications may not demand the Army’s extended storage periods, but manufacturers still need to establish what customers require after a battery has spent time in inventory. How long might it remain there? What storage temperatures and charge levels should testing represent? What performance must remain when the battery enters service?
Those questions belong in the purchasing specification and validation plan. Engineering and procurement should review storage evidence together before selecting a supplier. A lower unit price may offer little advantage if the cell cannot meet the customer’s requirements after an extended period in inventory.
Making Earlier Evidence Count
Manufacturers will still have to make decisions before the full lifetime data arrives, but early predictions can help teams investigate potential problems during a pilot while there is still time to change course. Advances in AI for materials discovery also allow researchers to explore more candidates and focus experiments on those with the strongest potential.
For manufacturers, these capabilities can provide better evidence about which materials to pursue and where to invest testing resources. Physical validation and production expertise remain essential to turning promising materials into reliable products.
The opportunity is to pursue longer-lasting, safer batteries with a clearer understanding of what to test next and what still needs to be proven before production.
Dr. Ang Xiao is the chief scientist of AI and quantum applications for batteries at SandboxAQ. The research discussed in this article is part of SandboxAQ’s work in battery performance prediction. The company has since expanded its battery AI programinto materials discovery, using Large Quantitative Models (LQMs) that combine machine learning with physics-based simulations to identify promising materials for next-generation batteries.
With more than 20 years of industry experience, Dr. Xiao’s expertise spans battery materials, PFAS, fluids, advanced analytical techniques, molecular simulations and machine learning. He led the technical strategy supporting a $500 million CHIPS R&D award focused on AI-driven domestic manufacturing of critical materials. He championed the development of lithium bis(trifluoromethanesulfonyl)imide (LiTFSI), an electrolyte salt also used as an additive and antistatic agent, with sales exceeding $200 million over the past decade. His work includes 40 papers and 17 patent applications, with more than 2,700 citations. He holds a Ph.D. in Chemistry from the University of Rhode Island and a Graduate Certificate in Data Science from Georgia Tech.
