It is now feasible to significantly expand the exploration of new materials.
It can require extended periods of meticulous effort — consulting information, conducting calculations, and performing precise laboratory tests — before scientists can introduce a new material with specific properties, whether it is for constructing an improved device or an enhanced battery. However, this process has been simplified thanks to artificial intelligence advancements.
A new AI algorithm named M3GNet, developed by researchers at the Jacobs School of Engineering at the University of California San Diego, is capable of predicting the structure and dynamic properties of any material, whether existing or newly conceptualized.
In fact, M3GNet was utilized to create a database containing over 31 million innovative materials that have not yet been synthesized, with their properties being forecasted by the machine learning algorithm. Moreover, this process occurs almost instantaneously.
Millions of potential materials
M3GNet can search for virtually any assigned material, such as metal, concrete, biological material, or any other type. To predict a material’s properties, the computer program needs to understand the material’s structure, which is based on the arrangement of its atoms.
In many ways, predicting new materials is similar to predicting protein structure — an area in which AlphaFold AI, developed by Google DeepMind, has excelled. Earlier this year, DeepMind announced the successful decoding of the structures of nearly all proteins in scientists’ catalogs, totaling over 200 million proteins.
Proteins, as the fundamental building blocks of life, undertake various functions within cells, from transmitting regulatory signals to protecting the body from bacteria and viruses. The accurate prediction of proteins’ 3D structures from their amino-acid sequences is therefore immensely beneficial to life sciences and medicine, and represents a revolutionary achievement.
Similar to how biologists previously struggled to decode only a few proteins over the course of a year due to inherent complexities, materials scientists can now invent and test novel materials much faster and more cost-effectively than ever before. These new materials and compounds can subsequently be integrated into batteries, drugs, and semiconductors.
“We need to know the structure of a material to predict its properties, similar to proteins,” explained Shyue Ping Ong, a nanoengineering professor at UC San Diego. “What we need is an AlphaFold for materials.”
Ong and his colleagues adopted a similar approach to AlphaFold, combining graph neural networks with many-body interactions to create a deep learning AI capable of scanning and generating practical combinations using all the elements of the periodic table. The model was trained using a vast database of thousands of materials, complete with data on energies, forces, and stresses for each material.
As a result, M3GNet analyzed numerous potential interatomic combinations to predict 31 million materials, over a million of which are expected to be stable. Additionally, the AI can conduct dynamic and complex simulations to further validate property predictions.
“For instance, we are often interested in how fast lithium ions diffuse in a lithium-ion battery electrode or electrolyte. The faster the diffusion, the more quickly you can charge or discharge a battery,” Ong stated. “We have demonstrated that the M3GNet IAP can accurately predict the lithium conductivity of a material. We firmly believe that the M3GNet architecture is a transformative tool that significantly enhances our ability to explore new material chemistries and structures.”
The Python code for M3GNet has been made available as open-source on Github for those interested. There are already plans to integrate this powerful predictive tool into commercial materials simulation software.
Google AI discovers 2.2 million new materials for various technologies
DeepMind has propelled researchers centuries ahead of the traditional pace of materials discovery methods.
Inorganic crystals play a crucial role in modern technologies. Their highly-ordered atomic structures provide them with unique chemical, electronic, magnetic, or optical properties that are utilized in a wide range of applications, from batteries to solar panels, microchips to superconductors.
Creating innovative inorganic crystals in a laboratory — whether to enhance an existing technology or to power a new one — is theoretically simple. A researcher sets up the conditions, conducts the procedure, and learns from failures to adjust conditions for the next attempt. This process is repeated until a new, stable material is obtained.
However, in practice, the process is exceedingly time-consuming. Traditional methods rely on trial-and-error guesswork that either modifies a known crystalline structure or makes speculative attempts. This process can be costly, time-intensive, and, if unsuccessful, may leave researchers with limited insights into the reasons for failure.
The Materials Project, an open-access database established at the Lawrence Berkeley National Laboratory, has revealed that about 20,000 inorganic crystals were discovered through human experimentation. Over the past ten years, researchers have utilized computational techniques to increase that number to 48,000.
Google’s AI research lab, DeepMind, has unveiled the results of a new deep-learning AI designed to forecast the potential structures of previously unidentified inorganic crystals. And the outcomes are far ahead of schedule.
DeepMind’s latest AI, named the “Graph Networks for Materials Exploration” (GNoME), is a graph neural network that identifies connections between data points through graphs.
GNoME was trained using data from the Materials Project and began creating theoretical crystal structures based on the 48,000 previously found inorganic crystals. It made predictions using two pipelines.
The first pipeline, known as the “structural pipeline,” relied on known crystal structures for its predictions. The second pipeline, known as the “compositional pipeline,” took a more randomized approach to explore potential molecule combinations.
The AI then validated its predictions using “density functional theory,” a method used in chemistry and physics to calculate atomic structures. Regardless of success or failure, this process generated more training data for the AI to learn from, informing future pipeline predictions.
In essence, the AI’s pipelines and subsequent learning mimic the human experimental approach mentioned earlier, but with the advantage of the AI’s processing power for faster calculations.
The researchers emphasize that, unlike language or vision, in materials science, data can continue to be generated, leading to the discovery of stable crystals, which can be used to further expand the model.
Overall, GNoME forecasted 2.2 million new materials, with around 380,000 considered the most stable and potential candidates for synthesis.
The potential inorganic crystals include layered, graphene-like that compounds could aid in the development of advanced superconductors and lithium-ion conductors that may enhance battery performance.
The authors of the study, Google DeepMind researchers Amil Merchant and Ekin Dogus Cubuk, stated that “GNoME’s discovery of 2.2 million materials would be equivalent to about 800 years’ worth of knowledge and demonstrate an unprecedented scale and level of accuracy in predictions.”
The research findings were published in the peer-reviewed journal Nature, and the 380,000 most stable materials will be freely available to researchers through the Materials Project, thanks to DeepMind’s contribution.
GNoME’s predicted materials are theoretically stable, but only 736 of them have been experimentally verified to date. This suggests that the model’s predictions are accurate to some extent but also highlights the long road ahead for experimental fabrication, testing, and application of all 380,000 materials.
To bridge this gap, the Lawrence Berkeley National Laboratory assigned their new A-Lab, an experimental lab that combines AI and robotics for fully autonomous research, to synthesize 58 of the predicted materials.
A-Lab, a closed-loop system, can make decisions without human input and has the capability to handle 50 to 100 times as many samples per day as a typical human researcher.
Yan Zeng, a staff scientist at A-Lab, noted that “We’ve adapted to a research environment, where we never know the outcome until the material is produced. The whole setup is adaptive, so it can handle the changing research environment as opposed to always doing the same thing.”
During a 17-day run, A-Lab successfully synthesized 41 of the 58 target materials, at a rate of more than two materials per day and with a success rate of 71%. The NBNL researchers published their findings in another Nature paper.
The researchers are investigating why the remaining 17 inorganic crystals did not materialize. In some cases, this may be attributed to inaccuracies in GNoME’s predictions.
In some cases, expanding A-Lab’s decision-making and active-learning algorithms might produce more favorable outcomes. Two instances involved successful synthesis after human intervention was attempted.
As a result, GNoME has provided A-Labs and human-operated research facilities worldwide with ample material to work with in the foreseeable future.
“My goal with the Materials Project was not only to make the data I generated freely available to expedite materials design globally, but also to educate the world on the capabilities of computations.
They can efficiently and rapidly explore vast spaces for new compounds and properties, surpassing the capabilities of experiments alone,” stated Kristin Persson, founder and director of the Materials Project.
He further stated, “In order to tackle global environmental and climate challenges, we need to develop new materials. Through materials innovation, we could potentially create recyclable plastics, utilize waste energy, manufacture better batteries, and produce more durable and affordable solar panels, among other things.”
In November, DeepMind, Google’s AI division, released a press statement titled “Millions of New Materials Discovered with Deep Learning.” However, researchers who analyzed a subset of the discoveries reported that they had not yet come across any strikingly novel compounds in that subset .
“AI tool GNoME finds 2.2 million new crystals, including 380,000 stable materials that could power future technologies,” Google announced regarding the discovery, stating that this was “equivalent to nearly 800 years’ worth of knowledge,” many of the discoveries “defied previous human chemical intuition,” and it represented “a tenfold increase in stable materials known to humanity.” The findings were published in Nature and garnered widespread attention in the media as a legacy to the tremendous potential of AI in science.
Another paper, simultaneously released by researchers at Lawrence Berkeley National Laboratory “in collaboration with Google DeepMind, demonstrates how our AI predictions can be utilized for autonomous material synthesis,” Google reported.
In this experiment, researchers established an “autonomous laboratory” (A-Lab) that utilized “computations, historical data from the literature, machine learning, and active learning to plan and interpret the outcomes of experiments conducted using robotics.
” Essentially, the researchers employed AI and robots to eliminate human involvement in the laboratory, and after 17 days, they had discovered and synthesized new materials, which they asserted “demonstrates the effectiveness of artificial intelligence-driven platforms for autonomous materials discovery.”
However, in the past month, two external groups of researchers who analyzed the DeepMind and Berkeley papers and published their own DNA at the very least suggest that this specific research is being overhyped.
All the materials science experts I spoke to emphasized the potential of AI in discovering new types of materials. However, they contend that Google and its deep learning techniques have not made a groundbreaking advancement in the field of materials science.
In a perspective paper published in Chemical Materials this week, Anthony Cheetham and Ram Seshadri from the University of California, Santa Barbara selected a random sample of the 380,000 proposed structures released by DeepMind and stated that none of them met a three-part test to determine whether the proposed material is “credible,” “useful,” and “novel.”
They argue that what DeepMind discovered are “crystalline inorganic compounds and should be described as such, rather than using the more general term ‘material,’” which they believe should be reserved for substances that “demonstrate some utility.”
In their analysis, they wrote, “We have not yet come across any surprisingly novel compounds in the GNoME and Stable Structure listings, although we anticipated that there must be some among the 384,870 compositions.
We also note that, while many of the new compositions are minor adaptations of known materials, the computational approach produces reliable overall compositions, which gives us confidence that the underlying approach is sound.”
In a phone interview, Cheetham informed me, “The Google paper falls short in terms of offering a useful, practical contribution to experimental materials scientists.” Seshadri stated, “We believe that Google has missed the mark with this.”
“If I were seeking a new material to perform a specific function, I wouldn’t sift through over 2 million proposed compositions as suggested by Google,” Cheetham mentioned. “I don’t think that’s the most effective approach.
I think the general methodology probably works quite well, but it needs to be much more targeted towards specific needs, as none of us have enough time in our lives to evaluate 2.2 million possibilities and determine their potential usefulness.”
We dedicated a significant amount of time to examining a small portion of the proposals, and we discovered that not only were they lacking in functionality, but most of them, although potentially credible, were not particularly innovative as they were essentially variations of existing concepts.
According to a statement from Google DeepMind, they stand by all the claims made in the GNoME paper.
The GNoME research by Google DeepMind introduces a significantly larger number of potential materials than previously known to science, and several of the materials predicted have already been synthesized independently by scientists worldwide.
In comparison to other machine learning models, the open-access material property database, The Materials Project, has acknowledged Google’s GNoMe database as top-tier. Google stated that some of the criticisms in the Chemical Materials analysis, such as the fact that many of the new materials have known structures but use different elements, were intentional design choices by DeepMind.
The Berkeley paper asserted that an “autonomous laboratory” (referred to as “A-Lab”) utilized structures proposed by another project called the Materials Project and employed a robot to synthesize them without human intervention, yielding 43 “novel compounds.” While a DeepMind researcher is listed as an author on this paper, Google did not actively conduct the experiment.
Upon analysis, human researchers identified several issues with this finding. The authors, including Leslie Schoop of Princeton University and Robert Palgrave of University College London, pointed out four common shortcomings in the analysis and concluded that no new materials had been discovered in that work.
Each of the four researchers I spoke to express that while they believe an AI-guided process for discovering new materials holds promise, the specific papers they evaluated were not necessarily groundbreaking and should not be portrayed as such.
“In the DeepMind paper, there are numerous instances of predicted materials that are clearly nonsensical. Not only to experts in the field, but even high school students could identify that compounds like H2O11 (a DeepMind prediction) do not seem plausible,” Palgrave conveyed to me.
“There are many other examples of clearly incorrect compounds, and Cheetham/Seshadri provide a more diplomatic breakdown of this. To me, it seems that basic quality control has been neglected—for the machine learning to output such compounds as predictions is concerning and indicates that something has gone wrong.”
AI has been employed to inundate the internet with vast amounts of content that is challenging for humans to sift through, making it difficult to identify high-quality human-generated work.
While not a perfect comparison, the researchers I consulted with suggested that a similar scenario could unfold in materials science: Having extensive databases of potential structures does not necessarily facilitate the creation of something with a positive societal impact.
“There is some value in knowing millions of materials (if accurate), but how does one navigate this space to find useful materials to create?” Palgrave questioned. “It is better to have knowledge of a few new compounds with exceptionally useful properties than a million where you have no idea which ones are beneficial.”
Schoop pointed out that there are already “50k unique crystalline inorganic compounds, but we only understand the properties of a fraction of these. So, it is unclear why we need millions more compounds if we have not yet comprehended all the ones we do know. It might be more beneficial to predict material properties than simply new materials.”
While Google DeepMind maintains its stand on the paper and disputes these interpretations, it is fair to say that there is now considerable debate regarding the use of AI and machine learning for discovering new materials, the context, testing, and implementation of these discoveries, and whether inundating the world with massive databases of proposed structures will truly lead to tangible breakthroughs for society, or simply generate a lot of noise.
“We do not believe that there is a fundamental issue with AI,” Seshadri remarked. “We think it is a matter of how it is utilized. We are not traditionalists who believe that these techniques have no place in our field.”
AI tools and advanced robotics are accelerating the search for urgently needed new materials.
In the latest development, researchers at Google DeepMind reported that a new AI model has identified over 2.2 million hypothetical materials.
Out of the millions of structures predicted by the AI, 381,000 were identified as stable new materials, making them prime candidates for scientists to fabricate and test in a laboratory.
Current Situation: The advancement of novel materials is crucial for the development of the next generations of the electrical grid, computing, and other technologies such as batteries, solar cells, and semiconductor chips.
This has led to significant investments in materials science and engineering by countries worldwide, including the US, China, and India. According to a report released this week from Georgetown’s Center for Security and Emerging Technology, AI and materials science are the top recipients of US federal grants to industry over the past six years.
China currently leads the field of materials engineering in several key areas, including publications, employment, and degrees awarded in the field.
Background: Traditionally, new materials were discovered by modifying the molecular structure of existing stable materials to create a novel material.
This method predicted approximately 48,000 stable inorganic crystal structures, with more than half of them being discovered in the past decade. However, this process is expensive, time-consuming, and less likely to yield radically different structures as it builds on known materials.
DeepMind accomplished this feat by utilizing existing data from the Materials Project at Lawrence Berkeley National Laboratory (LBNL) and other databases to train the AI, which then expanded the dataset as it learned.
How it Works: DeepMind’s tool, known as Graph Networks for Materials Exploration (GNoME), employs two deep learning models that represent the atoms and bonds in a molecule as a graph.
One of the models starts with known crystal structures and substitutes elements to create candidate structures. The other model, aiming to go beyond known materials and generate more diverse materials, uses only the chemical formula or composition of a candidate material to predict its stability.
The pool of candidates is filtered, and the stability of each structure is determined through energy measurements.
The most promising structures undergo evaluation with quantum mechanics simulations, and the resulting data is fed back into the model in a training loop known as active learning.
The Intrigue: GNoME seemed to grasp certain aspects of quantum mechanics and made predictions about structures it had not encountered.
For instance, despite being trained on crystals consisting of up to four elements, the AI system was able to discover five- and six-element materials, which have been challenging for human scientists, as stated by Ekin Dogus Cubuk, who leads the materials discovery team at DeepMind, during a press briefing.
The ability of an AI to generalize beyond its training data is significant. Keith Butler, a professor of computational chemistry materials at University College London, who was not involved in the research, emphasized the importance of exploring uncharted territories if these models are to be used for discovery.
What’s Next: Although predicting the stability of a potential structure does not guarantee its manufacturability.
In another paper published this week, researchers at LBNL shared the results from a lab equipped with AI-guided robotics for autonomous crystal synthesis. The material synthesis recipes were suggested by AI models that utilized natural language processing to analyze existing scientific papers and were then refined as the AI system learned from its mistakes.
Over a span of 17 days, operating 24/7, the A-Lab successfully synthesized 41 out of 58 materials they attempted to create, averaging more than two materials per day. However, some routes of synthesis may be challenging or costly to automate due to involving intricate glassware, movement across a lab, or other complex steps, as pointed out by Butler.
The A-Lab’s failures in synthesizing 17 materials were attributed to factors such as the requirement for higher heating temperatures or the need for better material grinding, which are standard procedures in a lab but fall outside the current capabilities of AI.
Ultimately, a material’s performance in conducting heat, being electronically insulating, or fulfilling other functions is essential. However, the synthesis and testing of a material’s properties are costly and time-consuming aspects of the process of developing a new material. Moreover, similar to many other AI systems, the models do not provide explanations for their decision-making process, as highlighted by Butler.
Butler also emphasized the impact of competitions to predict new structures and the influence of large language models (LLM) such as ChatGPT and other generative AI in the field.
The University of Liverpool’s Materials Innovation Factory, the Acceleration Consortium at the University of Toronto, and other organizations are working on developing self-driving laboratories.
According to Olexandr Isayev, a chemistry professor at Carnegie Mellon University involved in their automated Cloud Lab, certain scientific experiments can be effectively automated with machine learning and AI, although not all of them.
He also mentioned that the future progress in the field of science will be driven by the combination of software and hardware.
The expansion of this open-access resource is crucial for scientists who are striving to create new materials for upcoming technologies. With the help of supercomputers and advanced simulations, researchers can avoid the time-consuming and often ineffective trial-and-error process that was previously necessary.
The Materials Project, an open-access database established at the Lawrence Berkeley National Laboratory (Berkeley Lab) of the Department of Energy in 2011, calculates the properties of both existing and predicted materials.
Researchers can concentrate on the development of promising materials for future technologies, such as lighter alloys to enhance car fuel efficiency, more effective solar cells for renewable energy, or faster transistors for the next generation of computers.
Google’s artificial intelligence laboratory, DeepMind, has contributed nearly 400,000 new compounds to the Materials Project, expanding the available information for researchers. This dataset includes details about the atomic arrangement (crystal structure) and the stability (formation energy) of materials.
The Materials Project has the capability to visually represent the atomic structure of various materials. For example, one of the new materials, Ba₆Nb₇O₂₁, was computed by GNoME and consists of barium (blue), niobium (white), and oxygen (green). This was acknowledged by the Materials Project at Berkeley Lab.
Kristin Persson, the founder and director of the Materials Project at Berkeley Lab and a professor at UC Berkeley, stated, “We need to develop new materials to address global environmental and climate challenges. Through innovation in materials, we have the potential to create recyclable plastics, utilize waste energy, produce better batteries, and manufacture more cost-effective solar panels with increased longevity, among other possibilities.”
The Role of GNoME in Material Discovery
Google DeepMind created a deep learning tool called Graph Networks for Materials Exploration, or GNoME, to generate new data. GNoME was trained using workflows and data that had been developed over a decade by the Materials Project, and the GNoME algorithm was refined through active learning.
Ultimately, researchers using GNoME generated 2.2 million crystal structures, including 380,000 that are being added to the Materials Project and are predicted to be stable, thus potentially valuable for future technologies. These recent findings from Google DeepMind were published in the journal Nature.
Some of the calculations from GNoME were utilized in conjunction with data from the Materials Project to test A-Lab, a facility at Berkeley Lab where artificial intelligence guides robots in creating new materials. The first paper from A-Lab, published in Nature, demonstrated that the autonomous lab can rapidly discover new materials with minimal human involvement.
During 17 days of independent operation, A-Lab successfully produced 41 new compounds out of 58 attempts – a rate of over two new materials per day. In comparison, it can take a human researcher months of trial and error to create a single new material , if they are able to achieve it at all.
To create the novel compounds forecasted by the Materials Project, A-Lab’s AI generated new formulas by analyzing scientific papers and using active learning to make adjustments. Data from the Materials Project and GNoME were utilized to assess the predicted stability of the materials.
Gerd Ceder, the principal investigator for A-Lab and a scientist at Berkeley Lab and UC Berkeley, stated, “We achieved an impressive 71% success rate, and we already have several methods to enhance it. We have demonstrated that combining theory and data with automation yields remarkable results. We can create and test materials more rapidly than ever before, and expanding the data points in the Materials Project allows us to make even more informed decisions.”
The Impact and Future of the Materials Project
The Materials Project stands as the most widely accessed open repository of information on inorganic materials globally. The database contains millions of properties on hundreds of thousands of structures and molecules, with data primarily processed at Berkeley Lab’s National Energy Research Science Computing Center.
Over 400,000 individuals are registered as users of the site, and on average, over four papers citing the Materials Project are published each day. The contribution from Google DeepMind represents the most significant addition of structure-stability data from a group since the inception of the Materials Project.
Ekin Dogus Cubuk, lead of Google DeepMind’s Materials Discovery team, expressed, “We anticipate that the GNoME project will propel research into inorganic crystals forward. External researchers have already verified over 736 of GNoME’s new materials through concurrent, independent physical experiments, demonstrating that our model’s discoveries can be realized in laboratories.”
This one-minute time-lapse illustrates how individuals across the globe utilize the Materials Project over a four-hour period. The data dashboard showcases a rolling one-hour window of worldwide Materials Project activity, encompassing the number of requests, the users’ country , and the most frequently queried material properties. Credit: Patrick Huck/Berkeley Lab
The Materials Project is currently processing the compounds from Google DeepMind and integrating them into the online database. The new data will be freely accessible to researchers and will also contribute to projects such as A-Lab that collaborate with the Materials Project.
“I’m thrilled that people are utilizing the work we’ve conducted to generate an unprecedented amount of materials information,” said Persson, who also serves as the director of Berkeley Lab’s Molecular Foundry. “This is precisely what I aimed to achieve with the Materials Project: not only to make the data I produced freely available to expedite materials design worldwide, but also to educate the world on the capabilities of computations. They can efficiently and explore rapidly large spaces for new compounds and properties more efficiently and rapidly than experiments alone.”
By pursuing promising leads from data in the Materials Project over the past decade, researchers have experimentally confirmed valuable properties in new materials across various domains. Some exhibit potential for use:
– in carbon capture (extracting carbon dioxide from the atmosphere)
– as photocatalysts (materials that accelerate chemical reactions in response to light and could be employed to break down pollutants or generate hydrogen)
– as thermoelectrics (materials that could aid in harnessing waste heat and converting it into electrical power)
– as transparent conductors (which could be beneficial in solar cells, touch screens, or LEDs)
However, identifying these potential materials is just one of many stages in addressing some of humanity’s significant technological challenges.
“Developing a material is not for the faint-hearted,” Persson remarked. “It takes a long time to transition a material from computation to commercialization. It must possess the right properties, function within devices, scale effectively, and offer the appropriate cost efficiency and performance. The objective of the Materials Project and facilities like A-Lab is to leverage data, enable data-driven exploration, and ultimately provide companies with more viable opportunities for success.”
The researchers at Google DeepMind claim they have increased the number of stable materials known by ten times. Some of these materials could have applications in various fields such as batteries and superconductors, if they can be successfully produced outside the laboratory.
The robotic chefs were busy in a crowded room filled with equipment, each one performing a specific task. One arm selected and mixed ingredients, another moved back and forth on a fixed track tending to the ovens, and a third carefully plated the dishes.
Gerbrand Ceder, a materials scientist at Lawrence Berkeley National Lab and UC Berkeley, heard in approval as a robotic arm delicately closed an empty plastic vial—a task that he particularly enjoyed observing. “These robots can work tirelessly all night,” Ceder remarked, giving two of his graduate students a wry smile.
Equipped with materials like nickel oxide and lithium carbonate, the A-Lab facility is designed to create new and intriguing materials, particularly those that could be valuable for future battery designs. The outcomes of the experiments can be unpredictable.
Even a human scientist often makes mistakes when trying out a new recipe for the first time. Similarly, the robots sometimes produce a fine powder, while other times the result is a melted sticky mess or everything evaporates, leaving nothing behind. “At that point , humans would have to decide what to do next,” Ceder explained.
The robots are programmed to do the same. They analyze the results of their experiments, adjust the recipe, and try again, and again, and again. “You give them some recipes in the morning, and when you come back home, you might find a beautifully made soufflé,” said materials scientist Kristin Persson, Ceder’s close collaborator at LBNL (and also his spouse). Or you might return to find a burnt mess. “But at least tomorrow, they will make a much better soufflé.”
Recently, the variety of materials available for Ceder’s robots has expanded significantly, thanks to an AI program developed by Google DeepMind. This software, called GNoME, was trained using data from the Materials Project, a freely accessible database of 150,000 known materials overseen by Persson .
Using this information, the AI system generated designs for 2.2 million new crystals, out of which 380,000 were predicted to be stable—unlikely to decompose or explode, making them the most feasible candidates for synthesis in a lab. This has expanded the range of known stable materials almost tenfold. In a paper published in Nature, the authors state that the next solid-state electrolyte, solar cell materials, or high-temperature superconductor could potentially be found within this expanded database.
The process of discovering these valuable materials starts with actually synthesizing them, which emphasizes the need to work quickly and through the night. In a recent series of experiments at LBNL, also published in Nature, Ceder’s autonomous lab successfully created 41 of the theorized materials over 17 days, helping to validate both the AI model and the lab’s robotic techniques.
When determining if a material can be synthesized, whether by human hands or robot arms, one of the initial questions to ask is whether it is stable. Typically, this means that its atoms are arranged in the lowest possible energy state. Otherwise, the crystal will naturally want to transform into something else.
For thousands of years, people have been steadily adding to the list of stable materials, initially by observing those found in nature or discovering them through basic chemical intuition or accidents. More recently, candidate materials have been designed using computers.
According to Persson, the issue lies in bias: Over time, the collective knowledge has come to favor certain familiar structures and elements. Materials scientists refer to this as the “Edison effect,” based on Thomas Edison’s rapid trial-and-error approach to finding a lightbulb filament, testing thousands of types of carbon before settling on a variety derived from bamboo.
It took another decade for a Hungarian group to develop tungsten. “He was limited by his knowledge,” Persson explained. “He was biased, he was convinced.”
The approach by DeepMind aims to overcome these biases. The team started with 69,000 materials from Persson’s database, which is freely accessible and supported by the US Department of Energy. This was a good starting point, as the database contains detailed energetic information necessary to understand why certain materials are stable and others are not.
However, this was not enough data to address what Google DeepMind researcher Ekin Dogus Cubuk calls a “philosophical contradiction” between machine learning and empirical science. Similar to Edison, AI struggles to generate genuinely new ideas beyond what it has already encountered.
“In physics, you never want to learn something you already know,” Cubuk stated. “You almost always want to make generalizations outside of the known domain”—whether that involves discovering a different class of battery material or a new theory of superconductivity.
GNoME uses active learning, where a graph neural network (GNN) nests the database to identify patterns in stable structures and minimize atomic bond energy in new structures across the periodic table. This process generates numerous potentially stable candidates.
These candidates are then verified and adjusted using density-functional theory (DFT), a quantum mechanics technique. The refined results are incorporated back into the training data, and the process is repeated.
Through multiple iterations, the approach was able to produce more complex structures than those in the original Materials Project data set, including some composed of five or six unique elements, surpassing the four-element cap of the training data.
However, DFT is only a theoretical validation, and the next step involves the actual synthesis of materials. Ceder’s team selected 58 crystals to create in the A-Lab, considering the lab’s capabilities and available precursors. The initial attempts failed, but after multiple adjustments , the A-Lab successfully produced 41 of the materials, or 71 percent.
Taylor Sparks, a materials scientist at the University of Utah, notes the potential of automation in materials synthesis but emphasizes the impracticality of using AI to propose thousands of new hypothetical materials and then pursuing them with automation. He stresses the importance of tailoring efforts to produce materials with specific properties rather than blindly generating a large number of materials.
While GNNs are increasingly used to generate new material ideas, there are concerns about the scalability of the synthesis approach. Sparks mentions that the challenge lies in whether the scaled synthesis matches the scale of the predictions, which he believes is currently far from reality.
Only a fraction of the 380,000 materials in the Nature paper are likely to be feasible for practical synthesis. Some involve radioactive or prohibitively expensive elements, while others require synthesis under extreme conditions that cannot be replicated in a lab.
This practicality challenge extends to materials with potential for applications such as photovoltaic cells or batteries. According to Persson, the bottleneck consistently lies in the production and testing of these materials, especially for those that have never been made before.
Furthermore, turning a basic crystal into a functional product is a lengthy process. For instance, predicting the energy and structure of a crystal can help understand the movement of lithium ions in an electrolyte, a critical aspect of battery performance. However, predicting the electrolyte’s interactions with other materials or its overall impact on the device is more challenging.
Despite these challenges, the expanded range of materials opens up new possibilities for synthesis and provides valuable data for future AI programs, according to Anatole von Lilienfeld, a materials scientist at the University of Toronto.
Additionally, the new materials generated by GNoME have piqued the interest of Google. Pushmeet Kohli, vice president of research at Google DeepMind, likes GNoME to AlphaFold and emphasizes the potential for exploring and expanding the new data to address fundamental problems and accelerate synthesis using AI .
Kohli stated that the company is considering different approaches to directly engage with physical materials, such as collaborating with external labs or maintaining academic partnerships. He also mentioned the possibility of establishing its own laboratory, referring to Isomorphic Labs, a spinoff from DeepMind focused on drug discovery, which was founded in 2021 after the success of AlphaFold.
Researchers may encounter challenges when attempting to apply the materials in practical settings. The Materials Project is popular among both academic institutions and businesses because it permits various types of use, including commercial activities.
Google DeepMind’s materials are made available under a distinct license that prohibits commercial usage. Kohli clarified, “It’s released for academic purposes. If individuals are interested in exploring commercial partnerships, we will evaluate them on a case-by-case basis.”
Several scientists working with new materials observed that it’s unclear what level of control the company would have if experimentation in an academic lab leads to a potential commercial application for a GNoME-generated material. Generating an idea for a new crystal without a specific purpose in mind is generally not eligible for a patent, and it could be challenging to trace its origin back to the database.
Kohli also mentioned that although the data is being released, there are currently no plans to release the GNoME model. He cited safety concerns, stating that the software could potentially be used to generate hazardous materials, and expressed uncertainty about Google DeepMind’s materials strategy. ” It’s difficult to predict the commercial impact,” Kohli said.
Sparks anticipates that his colleagues in academia will be displeased with the absence of code for GNoME, similar to biologists’ reaction when AlphaFold was initially published without a complete model. “That’s disappointing,” he remarked.
Other materials scientists are likely to want to replicate the findings and explore ways to enhance the model or customize it for specific purposes. However, without access to the model, they are unable to do so, according to Sparks.
In the interim, the researchers at Google DeepMind hope that the discovery of hundreds of thousands of new materials will be sufficient to keep both human and robotic theorists and synthesizers occupied. “Every technology could benefit from improved materials. It’s a bottleneck,” Cubuk remarked “This is why we need to facilitate the field by discovering more materials and assisting people in discovering even more.”
