5 AI Breakthroughs Transforming Materials Discovery Today

Artificial intelligence (AI) continues to transform the world and reshape the future of humanity.

The technology is driving changes across nearly every sector by performing tasks that typically require human intelligence. AI systems utilize vast amounts of data to identify patterns and make decisions. 

This way, AI is able to simulate certain levels of human-like reasoning and cognitive processes.

According to the World Trade Report, AI’s productivity gains and cost benefits can boost global GDP by 12-13% by 2040.

By narrowing their digital infrastructure gap with high-income economies by 50% and adopting AI more widely, low- and middle-income economies can see as much as 15% increase in their incomes.

Besides helping nations bolster their productivity, trade, and economic position, AI can help society by driving innovations across industries. One of the ways the technology is currently doing that is through materials discovery. 

The Promise of AI in Material Discovery

The discovery of materials has always been key to innovation. Many centuries ago, the mix of copper and tin led to the Bronze Age, when stronger tools and weapons transformed trade and societies.

Then came the Iron Age, when mastery of iron reshaped economies. Fast forward to the 19th century, and steel gained widespread adoption. An alloy of iron and carbon, steel was the backbone of railways, skyscrapers, ships, and machinery, fueling the Industrial Revolution and global expansion.

In the late 20th century, the Silicon Age transformed the world with the discovery and refinement of semiconductors that are the foundation of modern electronics. We are now in the era of advanced materials, where graphene, carbon nanotubes, and quantum materials are opening doors to cleaner energy, lighter aircraft, and faster computing.

The advent of AI and machine learning (ML) is contributing to innovation in materials and, by extension, various industries by significantly accelerating the process of material discovery, design, and optimization.

For this, AI utilizes algorithms and models to screen vast databases of candidates for specific application needs. Here, deep learning models like Graph Neural Networks (GNNs) and Recurrent Neural Networks (RNNs) are crucial for analyzing the complex datasets that are found in materials science.

They can also identify existing materials with desired properties from these databases and even predict the properties of materials based on their composition and structure. 

With the help of AI, the field of material science can move beyond the traditional trial-and-error methods, which are time-consuming and expensive.

Moreover, AI models can generate novel material structures tailored to specific requirements. When integrated with automated experimental platforms, AI can speed up the long process of material discovery to production.

Despite these benefits, challenges remain in terms of the lack of quality and extensive data for certain materials. Successful synthesis of newly found and designed materials in the laboratory is another big challenge.

As materials scientist Anthony Cheetham from UCSB remarked¹ in Nature after examining the list of 2.2 million hypothetical crystals found by GNoME, an AI tool by DeepMind, a subsidiary of Alphabet (Google), “It’s one thing to discover a compound, and a totally different thing to discover a new functional material.”

Further noting the impracticality of many AI-predicted compounds, Cheetham said:

“We found quite a lot of things that were ridiculous.”

This shows the gap between prediction and practical realization. What this gap requires is the combination of AI with human expertise and experimental science.

Still, the promise of AI to revolutionize materials science simply can’t be dismissed. Given its ability to lead to quicker development of materials for energy, healthcare, automotive, aerospace, and other crucial applications, that impact is too big to ignore.

So, let’s take a look at some of the most prominent examples of AI’s application in material science that showcase its potential to push the boundaries of material discovery and innovation. 

Swipe to scroll →

1. Perovskite Solar Cells: AI-Optimized Materials and Processing

One of the most promising solutions to achieve sustainable energy is solar power, and its adoption is rapidly increasing. In 2024, the world installed a record ~600 GW of solar power, rising 33% from 2023. By the end of the decade, this is expected to reach ~1 TW per year.

This growing demand for solar energy creates the need for more efficient, versatile, and cost-effective materials for solar cells.

Perovskite is one such material that offers a unique crystalline structure. The naturally occurring mineral can now be recreated synthetically. By mixing organic and inorganic elements, scientists create synthetic perovskites that exhibit remarkable light-absorbing properties, making them highly suitable for solar applications.

Besides high efficiency, these materials offer the benefits of flexibility and tunable bandgap, but scalability and stability issues persist; hence, the search for new compositions.

So, researchers turned to AI for correlating the performance of perovskite solar cells (PSCs) with material properties and energy conversion processes more than a decade ago. They then utilized the technology to optimize material composition, develop design strategies, and predict performance.

In 2019, researchers from the University of Central Florida reviewed² over 2,000 peer-reviewed publications about perovskite to collect over 200 data points, which were fed to the AI system they created to get the best recipe for perovskite solar cells (PSC). The same year, scientists at MIT developed³ a model to speed up the synthesis and analysis of new compounds by a factor of ten and found two new lead-free perovskites worthy of further investigation.

In 2022, researchers from MIT and Stanford University reported⁴ taking the help of AI to scale up advanced solar cell manufacturing.

For this, a system was created, which has been in development for several years, to integrate data from prior experiments as well as information based on experienced workers’ personal observations. This integration made outcomes more accurate and led to the manufacturing of perovskite cells with an energy conversion efficiency of 18.5%.

This is unlike most machine-learning systems, which primarily use raw data and typically do not incorporate human experience. To include outside information in their model, they used a probability factor based on Bayesian Optimization, allowing them to “find out trends that we weren’t able to see before.”

The discovery of advanced perovskite solar technology with the help of AI is continuing and gaining speed to increase PSC efficiency. In one such study⁵, the efficiency was increased to 26.2% while saving “enormous amounts of time and resources.”

⬆ Back to table

2. AI-Discovered Electrocatalysts for Hydrogen Production

A promising substitute for non-renewable fossil fuels that are responsible for tremendous amounts of greenhouse gas (GHG) emissions is hydrogen. The most abundant element in the universe, hydrogen, has emerged as a clean and renewable energy source.

However, efficient production of hydrogen to meet its commercial-scale demand is a serious challenge. Here, water splitting electrolysis offers a promising pathway, where electrocatalysis plays a critical role. This makes the development of low-cost, active, and stable electrocatalysts an essential prerequisite for getting desired electrocatalytic hydrogen production from water splitting.

Electrocatalysts accelerate hydrogen production by lowering the energy needed for water splitting by utilizing expensive precious metals like platinum or more affordable alternatives like nickel, cobalt, graphene, MXenes, and others.

Besides the properties and cost of the material, a specific catalyst is chosen based on whether the reaction is acidic, alkaline, or operates at high temperatures. 

However, it is very time-consuming and expensive to use the traditional trial and error method to search existing and new suitable materials to improve the reactions, so AI is being utilized⁶ to overcome the limitations of traditional approaches, discover novel candidates, and improve known products.

A recent study reported⁷ that its entropy-screened AI trained on a DoE dataset looked through 16.2 million chemical compositions to identify Fe12Co28Ni33Mo17Pd5Pt5 as the best composition for water splitting. The alloy shows ultralow overpotentials for both fundamental electrocatalytic reactions, HER and OER, while having robust stability.

Meanwhile, a couple of years ago, Google AI lab DeepMind contributed 380,000 new compounds to the Materials Project, a platform that underpins many catalyst searches and autonomous experiments. 

The open-access database founded at the Department of Energy’s Berkeley Lab has been used by researchers to experimentally confirm useful properties in new materials that show potential for use in carbon capture and as photocatalysts, thermoelectrics, and transparent conductors.

The database includes how the atoms of a material are arranged and how stable it is. GNoME was trained using the data and workflows that were developed by the Project and then improved via active learning.

Using the computations from Google DeepMind’s GNoME along with data from the Materials Project, the researchers tested A-Lab, a facility at Berkeley Lab where AI guides robots in making new materials. The A-Lab successfully produced⁸ 41 new compounds.

⬆ Back to table

3. Superhard Materials: ML-Guided Discovery Beyond Diamond

Industries like military, aerospace, and energy production demand superhard materials, which are virtually incompressible solids. The hardness value of these materials surpasses 40 gigapascals (GPa) on the Vickers scale, and they have high bond covalency and high electron density.

Diamond is the hardest known material to date, boasting a hardness value in the range of 70-150 GPa. What this means is that it would take more than this much (70-150 GPa) pressure to leave an indentation on the diamond’s surface. As a result, it is used in cutting tools, abrasives, wear-resistant coatings, and for creating high-pressure experiments.

These precious stones, which are a solid form of the element carbon with its atoms arranged in a diamond cubic crystal structure, are also used by scientists to find new suitable materials. But AI has changed that.

Over the years, several researchers have found⁹ new superhard phases, with one reporting¹⁰ BC10N, B4C5N3, and B2C3N to exhibit dynamically stable phases with hardness values >40 GPa.

In 2020, researchers from the University of Houston and Manhattan College utilized¹¹ an ML model to accurately predict the hardness of new materials, thus enabling them to find appropriate compounds more readily.

The amount of high pressure required to make any mark on a material’s surface makes them rare, and “identifying new materials challenging.” And this is exactly why, “materials like synthetic diamond are still used even though they are challenging and expensive to make,” said paper’s co-author Jakoah Brgoch, who’s an associate professor of chemistry at the University of Houston.

A complicating factor here is load dependence, which means that a material’s hardness may vary depending on the amount of pressure exerted. This makes testing a material experimentally complex. Even using computational modeling is almost impossible, so the researchers created a model that overcomes the challenge by predicting the load-dependent Vickers hardness based only on the material’s chemical composition.

The algorithm was trading on a database that involved 560 different compounds that required going through hundreds of academic papers. “All good machine learning projects start with a good dataset,” said Brgoch. “The true success is largely the development of this dataset.”

As a result, they found over 10 new stable borocarbide phases, and with the model’s accuracy at 97%, they feel hopeful about achieving success in the lab.

AI isn’t without its limitations, though, as Brgoch noted, “The idea of using machine learning isn’t to say, ‘Here is the next greatest material,’ but to help guide our experimental search.” What the technology does is “it tells you where you should look.”

⬆ Back to table

4. Polymer Dielectrics: AI-Accelerated Energy-Storage Materials

An essential component of modern energy storage is dielectrics, which are non-conductive materials such as air, glass, and plastic. 

The choice of dielectric material is what determines the energy density of capacitors, and polymer dielectrics are widely utilized for energy storage because of their low cost, mechanical flexibility, reliability, fast discharge speed, and ease of processing. But again, their low energy density is a problem. 

As a result, researchers continually look to improve performance by developing new polymer dielectrics to increase their energy storage capacity for applications in power systems, electronics, and electric vehicles (EVs). 

AI has made amazing progress in polymer materials. For instance, just a couple of months ago, researchers at MIT and Duke University collaborated to create¹² more durable polymers by incorporating stress-responsive crosslinker molecules, which were identified by AI. MIT researchers have also built¹³ a system that finds, mixes, and tests as many as 700 new polymer blends a day for applications like battery electrolytes, protein stabilization, or drug-delivery materials.

Designing new polymer blends presents the problem of an almost endless number of possible polymers to start with, and once a few have been selected to mix, then the composition of each polymer must be chosen, as well as the concentration of polymers in the blend.

“Having that large of a design space necessitates algorithmic solutions and higher-throughput workflows because you simply couldn’t test all the combinations using brute force.”

– Paper’s senior author, Connor Coley

Their AI system provided them with optimal blends, with the best one performing 18% better than its individual components.

Given the efficiency with which AI provides new polymer options and blends, it makes sense to apply the technology¹⁴ to identify better polymer dielectrics¹⁵.

A team of researchers did just that and discovered¹⁶ dielectrics with 11 times the energy density of commercial alternatives at elevated temperatures.

The innovative algorithm is developed to predict the properties and formulations of polymers before actually creating them. For this, they first defined specific requirements and then trained the ML models on existing material-property data to predict desired outcomes.

Besides AI, the researchers employed established polymer chemistry and molecular engineering to discover a suite of dielectrics in the polynorbornene and polyimide families, with many of the discovered dielectrics displaying high energy density and high thermal stability over a broad temperature range. 

But one in particular exhibited an energy density of 8.3 J cc−1 at 200 °C, which is much higher than the polymer dielectric commercially available.

“In the early days of AI in materials science, propelled by the White House’s Materials Genome Initiative over a decade ago, research in this field was largely curiosity-driven. Only in recent years have we begun to see tangible, real-world success stories in AI-driven accelerated polymer discovery,” said co-author Rampi Ramprasad, a Professor at Georgia Institute of Technology. “These successes are now inspiring significant transformations in the industrial materials R&D landscape.”

⬆ Back to table

5. Solid Electrolytes: AI for Safer, Higher-Density Batteries

Driven by the widespread adoption of portable devices and EVs and the increasing demand for renewable energy storage solutions, the global battery market is rapidly advancing¹⁷. Given the important role batteries play in the modern world, scientists are constantly trying to develop more energy-efficient and safer battery technology.

While lithium-ion batteries are most widely used today, they have limited lifespan and safety risks, which are addressed by solid-state batteries (SSBs). 

These batteries replace liquid electrolytes with solid-state electrolytes to eliminate the risk of inflammation at elevated temperatures while enabling higher energy density and improving durability, creating safer and more powerful batteries. 

But these batteries with solid electrolytes face their own challenges, such as low ionic conductivity, electrode interface compatibility, mechanical and chemical stability, and cost-effective manufacturing. So, researchers are exploring materials that can overcome these issues through AI.

Unlike other fields we discussed today, batteries are one of the hottest areas where the application of AI¹⁸ has exploded due to the involvement of leading carmakers and startups that are pouring money into solid-state battery R&D. Besides the safety risk, the sector has also accumulated big databases, which are rich enough to train ML models.

Even governments have listed SSBs as a strategic priority to secure domestic supply chains and achieve national energy and climate goals.

So, there are various instances where AI has helped¹⁹ researchers and companies discover new solid electrolytes.

Last year, Microsoft researchers used AI along with supercomputers to sift through 32 million potential inorganic materials to find 18 promising candidates²⁰ in a matter of a few days. The new material, N2116, is a solid-state electrolyte that can reduce lithium use in batteries by 70% and has been tested to power a lightbulb.

DeepMind’s AI tool GNoME, meanwhile, has identified²¹ 528 promising lithium-ion conductors, some of which may help make batteries more efficient. 

Then there’s LBS²² (Li8B10S19) from Stanford researchers, who called it “the most stable, sulfur-based lithium-ion electrolyte that we’ve ever seen experimentally.” The researchers first identified²³ solid electrolytes to someday replace flammable liquid electrolytes in Li-ion batteries via AI about a decade ago.

⬆ Back to table

Conclusion