AI Discovers New Materials for Next-Gen Batteries

Lithium Versus The Rest

Lithium-ion batteries have so far dominated the electrification landscape, in large part due to lithium atoms’ unique electrical properties. Simply put, lithium, being the 3^rd lightest element in the periodic table, is the most powerful one when it comes to carrying charges with a single electron.

Source: Medium

However, lithium is expensive, making alternative battery chemistries potentially economically attractive. Notably, sodium-ion batteries have been gaining traction for this very reason.

It appears that another design might have more potential than previously thought: multivalent-ion batteries. They used metallic ions that can carry more than one electron at once, and might be more cost-effective than lithium-ion batteries.

The recent breakthrough was achieved by using AI to test millions of combinations for the battery materials. This discovery was done by researchers at the New Jersey Institute of Technology (NJIT) and the  Rensselaer Polytechnic Institute (RPI) (USA). They published their results in Cell Reports Physical Science¹, under the title ”Generative AI for discovering porous oxide materials for next-generation energy storage”.

The Many Types of Ion Batteries

If lithium-ion took over small electronics and early EV designs thanks to its energy density, many other metallic ions can be used according to the same principle.

As discussed, sodium-ion is currently a popular alternative, increasingly mass-manufactured for cheap EV models.

Another option is to use magnesium, calcium, aluminum, or zinc, all of which are multivalent ions. This means that they carry two or even three positive charges.

However, the larger charge also comes with larger sizes. Both larger atomic specs make multivalent ions challenging to accommodate efficiently in battery materials, reducing the battery density too much to be commercially viable.

At least, this was true with conventional battery materials developed for lithium or sodium ions. But they are far from the only possible battery materials that can be used. Many other crystalline structures could be built to harbor the ions whose movement carries the electric charges.

“One of the biggest hurdles wasn’t a lack of promising battery chemistries — it was the sheer impossibility of testing millions of material combinations,”

Professor Dibakar Datta – New Jersey Institute of Technology (NJIT)

AI Assisting Research

A Powerful Assistant

Human minds are not the best at handling any data set where the numbers are going toward the millions. But AIs are excellent at it.

It is a growing trend of researchers, especially in material sciences or biotech, using AI technology to help identify the most promising ideas, before analyzing and testing them more rigorously.

“We turned to generative AI as a fast, systematic way to sift through that vast landscape and spot the few structures that could truly make multivalent batteries practical.

Professor Dibakar Datta – New Jersey Institute of Technology (NJIT)

Previously, a computing model relying solely on physics would have been unable to handle the extremely complex calculation required to model a new type of crystal structure.

But new types of AI, based on machine learning and neural networks, are more able to “guess” the general property of a material without formal mathematical calculation of the physics behind.

Source: Cell

The researchers developed a system relying on two different types of AI at the same time, one knowledgeable about crystals, and one LLM (Large Language Model), the same technological base behind ChatGPT.

Source: Cell

Crystal Diffusion Variational Autoencoder (CDVAE)

The CDVAE model generated 10,000 structures, which were subjected to a series of precise screening and validation steps to ensure they met the necessary standards.

For example, it checked that the distance between atom pairs was large enough, or the charge neutrality of the system.

This method generated 42 structures potentially usable for battery materials.

Of these, 21 structures matched existing entries in the database but offered new configurations with differences in stoichiometry, lattice parameters, or space groups. The remaining 21 structures were entirely novel.

Source: Cell

So it created both new versions of existing material that were previously unknown and have entirely new potential battery material.

LLM

The researchers then used Meta’s (FB +0.39%) Llama-3.1-8B, specially calibrated and tailored for generating crystal structures.

Source: Cell

This generated 10,000+ crystal structures, of which 1,087 structures remained after checking for structural integrity. Using the same filters as with CDVAE, this resulted in 13 potential candidates, of which the 5 most stable structures were chosen.

Source: Cell

Swipe to scroll →

Challenging The AIs’ Findings

The researchers used a mathematical testing method called “DFT relaxation”, calculating the material free energy (linked to stability), to check the quality of the material found.

It quickly appears that the LLM-generated crystalline materials were generally much better and more stable than those generated with CDVAE.

Source: Cell

“Our AI tools dramatically accelerated the discovery process, which uncovered five entirely new porous transition metal oxide structures that show remarkable promise,”

Professor Dibakar Datta – New Jersey Institute of Technology (NJIT)

Can It Be Made?

The team validated their AI-generated structures using quantum mechanical simulations and stability tests, confirming that the materials could indeed be synthesized experimentally and hold great potential for real-world applications.

“These materials have large, open channels ideal for moving these bulky multivalent ions quickly and safely, a critical breakthrough for next-generation batteries.”

Professor Dibakar Datta – New Jersey Institute of Technology (NJIT)

The next step will be to collaborate with experimental labs to synthesize and test the newly conceptualized AI-designed materials.

It could make multivalent batteries the next step in battery technology. So far, the lack of a good material to accept the larger atoms has blocked the development of this option. By using better materials to store magnesium, aluminum, or other large ions, the multi-electron transport capacity of these atoms could maybe one day outshine even lithium’s powerful, but single-electron power-carrying capacity.

Investing in Material Science & AI Innovation

Meta: AI-Driven Material Science

Today, Meta is still mostly known for its Facebook and Instagram social media platforms, as well as WhatsApp chat. It is also present in the virtual reality (VR) space with its VR Headsets and its somewhat unsuccessful “Metaverse”.

Importantly, though, Meta is an AI company with massive investments in infrastructure to make it happen.

“The first multi-gigawatt data center, dubbed Prometheus, is expected to come online in 2026, while another, called Hyperion, will be able to scale up to 5 gigawatts over the coming years.

“We’re building multiple more titan clusters as well. Just one of these covers a significant part of the footprint of Manhattan.

Mark Zuckerberg – Meta Founder & CEO

LLM technology seems at first glance to be mostly useful for “talking” tasks, like chatbots, improved online search, education, and other human-centric activities.

But this research illustrates that the ability of LLM to learn language can be deployed to other data-heavy tasks, like learning how to “talk” crystal structures. The same can be said for genetic codes, for example.

This means that progress in LLM algorithms will likely usher in a golden age of entirely new discoveries in the creation of new materials for batteries, advanced materials, energy generation, etc., as well as new types of proteins and DNA/RNA materials that can be turned into medicine, biomanufacturing tools, etc.

In that context, this implies that companies like Meta and its Lama LLM models are not only building potentially profitable tech replacements for existing tools, but also could turn into an IP powerhouse in the physical world as well.

In that context, it might be remembered that the original tech business of companies like Meta, or for that matter Google (GOOGL +0.17%) or Microsoft (MSFT -0.46%), was just a stepping stone before turning them into AI & IP-driven giants changing the world with many new technologies, including in renewable energy and material sciences.

Latest Meta (META) Stock News and Developments

Study Referenced

^{1. Joy Datta, Amruth Nadimpally, Nikhil Koratkar, Dibakar Datta. Generative AI for discovering porous oxide materials for next-generation energy storage. Cell Reports Physical Science, Volume 6, Issue 7, 102665. July 16, 2025. https://www.cell.com/cell-reports-physical-science/fulltext/S2666-3864(25)00264-4}