How Is AI Accelerating the Discovery of New Materials?
Will the next great carbon capture material or battery design come from the mind of AI?

Complete the form below to unlock access to ALL audio articles.
Batteries, solar panels, computer chips, carbon capture systems. All these innovative technologies, and others like them, are the result of serious breakthroughs in materials science – driven by the discovery and synthesis of novel inorganic materials.
For decades, the discovery of new inorganic materials with more favorable properties was a daunting task of trial and error, with scientists forced to conduct hundreds upon hundreds of hours of painstaking experimentation to identify and synthesize just a handful of potential new materials.
Computational chemistry was a revolution in the world of materials science when it was first introduced. With the increased availability of supercomputers and the combined efforts of physicists, chemists and computer scientists, researchers could simulate the behavior of molecules and materials at the atomic scale. This helped scientists to accurately predict the properties of new materials without the need for such repetitive physical experimentation, shedding a significant amount of the “trial and error” baggage.
Today, with the advent of machine learning (ML) and artificial intelligence (AI), there is a sense that another materials revolution could be on the way, with AI-guided materials discovery set to accelerate these computational approaches even further.
Better materials for tackling the climate crisis
“Let me just say that I am not a material scientist, I am a theoretical physicist,” Dr. Eliu A. Huerta began our conversation. Huerta is the lead for translational AI in the Data Science and Learning Division at Argonne National Laboratory, US, and has been working at the intersection of AI and scientific research for a decade already. Under his guidance, researchers at Argonne have been applying AI and advanced computational techniques to tackle grand challenges in astrophysics, cosmology, materials science and biophysics.
“One thing I learned when I was being trained as a theoretical physicist is the value of being an outsider. You look at things in a completely different way and this allows you to propose new ideas,” Huerta told Technology Networks.
“I am really excited about solving problems that AI alone cannot solve, that domain knowledge alone cannot solve, that supercomputing alone cannot solve — but where a mix of all of these can provide new approaches and new opportunities to understand science in a way you couldn’t do with these separate tools,” said Huerta.
Huerta recently helped lead a project that sought to design new materials for carbon capture. Specifically, the group was interested in a class of compounds known as metal-organic frameworks (MOFs). MOFs are highly porous materials, made up of metal ion clusters and organic ligands which function as the network’s nodes and linkers respectively.
Published in Communications Chemistry, the researchers used a generative AI diffusion model to suggest unique and chemically diverse linkers that could be used to make novel MOFs. These were screened using a modified neural network that would select the MOFs with the best theoretical carbon capture performance. The final structures were then validated using traditional computational chemistry methods, including molecular dynamics and grand canonical Monte Carlo simulations, to compute more credible CO2 absorption capabilities and make the final selection of best performers.
“Typically, when you do computational chemistry, you know the structure that you want to validate. You know that the molecule already exists and you want to go and measure some properties. But we were interested in doing this and borrowing ideas from drug design and discovery,” Huerta explained. “So here we use the diffusion model, which is a generative AI model, and we expose that model to different molecular structures so that the model would learn not only about metal-organic frameworks, but about physics broadly speaking. I think this was the key to allowing the model to propose some chemical structures that were entirely novel."
The strength of this kind of approach is its speed, Huerta continued. Creating new MOFs has been the focus of researchers for several decades, yet the number of exceptionally high-performing MOF materials is still relatively low.
“Experimental science is needed, but it maybe is not the optimal way to go and discover new materials,” Huerta said. “Then again, if you are only using computational chemistry methods, it is very challenging to create a new material from the ground up! Now, with the method that we are proposing, we are learning from experimental chemistry and computational chemistry, but we are allowing AI to go and explore this vast chemical design space and find new things that we did not know about in the past.”
In the study, Huerta’s team was able to generate over 120,000 MOF candidates in 33 minutes using a supercomputer at the Argonne Leadership Computing Facility. This was whittled down by the modified neural network to 364 AI-generated MOFs that were believed to be high-performing. In total, this process took just over five hours. Further computational analysis, which took only a few days to complete, found 102 stable MOFs in this dataset, of which 6 had a CO2 capacity that ranked in the top 5% of materials in the popular hMOF database.
With the team’s tool having found success in suggesting novel, high-performance carbon capture materials, Huerta believes that similar workflows with altered parameters could help with accelerating the design of advanced materials for other applications.
“When you have these tools, you want to develop the ability to go and tackle similar problems that you can apply MOFs in, for example, hydrogen storage. This is just another parameter that you can use to fine-tune your generative AI model. Going beyond this, there is also methane capture, with methane being another gas that is responsible for environmental pollution,” Huerta said. “There are many applications where we can use this software to go and explore the properties of new materials.”
Improving batteries with AI-generated materials
Machine learning and generative AI are useful for more than just MOFs. As tools, they can be applied in a similar way to accelerate the design and discovery of other classes of material for different applications.
“Each generation faces a defining technical challenge, and for ours, that challenge is climate change. It’s urgent and requires immediate action,” Dr. Austin Sendek, an adjunct professor of materials science & engineering at Stanford University, told Technology Networks.
“AI has the potential to accelerate fundamental scientific processes. My focus on energy technologies is driven by both this scientific potential and the broader global impact of solving this issue,” Sendek said.
At Stanford, Sendek’s research focuses on harnessing the power of machine learning and AI to design new materials that can support the decarbonization of the global economy. The main focus of this research is on batteries and electrochemistry – using machine learning to assist in the discovery of better electrolytes to support high-performance batteries.
“Electrochemistry offers a new modality through which we can achieve many of the same outcomes as burning fuels, but by using clean, renewable electrons instead,” Sendek explained. “While powering vehicles through electricity via batteries is a major application, the scope of next-generation electrochemistry extends to areas like cement production, grid power backup and the creation of various materials, for instance.”
The search for new and improved battery materials and electrolytes is a core part of the decarbonization effort. Higher-performance batteries can help store more energy from the grid when renewable energy production is high, meaning that more green energy can still be released even when production is low. Battery improvements could also lead to better electric vehicles with larger ranges or faster charging times – again helping to support the transition away from fossil fuels.
In a 2018 paper, published in the journal Chemistry of Materials, Sendek and co-authors demonstrated the use of a machine learning-based prediction model to generate novel lithium ion conductors for use in all-solid-state batteries.
“Electrolyte components heavily impact the performance and properties of these cells,” Sendek noted. “There’s a vast chemical space to explore, with over 10 billion commercially available molecules that can be used to modify electrolytes in various ways.”
The team found that this machine learning-assisted approach was 2.7 times more likely to identify fast lithium conductors than a random search, in addition to performing well in a head-to-head competition against six PhD students with experience in the field.
In a review published in Advanced Energy Materials, Sendek and his colleagues also highlight the ability of machine learning-based approaches to assist in other areas, such as process optimization, cell lifetime prediction and battery modeling, in addition to accelerating materials discovery.
To further explain why AI and machine learning are such helpful tools, Sendek strips it back to a matter of depth and breadth.
“On the depth side, AI and machine learning help us identify patterns and relationships that would typically require scientific intuition and principles to uncover. For example, it can reveal how different electrolytes affect battery performance or predict the properties of new materials — tasks that, in the past, might have taken years of experimentation,” he said.
“On the breadth side, once we develop predictive models, machine learning allows us to apply them to a much larger design space at incredible speeds. These models can rapidly evaluate vast numbers of molecules, often identifying potential candidates that would have taken traditional scientific methods much longer to find.”
The future outlook for materials and AI
AI and machine learning are quickly solidifying themselves as novel, useful tools in many areas of science – including medicine, proteomics, drug discovery and more. In materials science, the benefit of these tools is their speed. With today’s computing power, these tools can generate realistic novel materials at a blistering pace, freeing scientists from extensive loops of “trial and error” and allowing them to put their expertise to better use further down the development process. In the context of the global climate crisis, the search for innovative new materials for carbon capture and green energy storage cannot happen quickly enough.
However, the application of AI and machine learning to materials science is still a relatively nascent field, with several limitations needing to be addressed as the field develops. For example, much of AI’s use in materials discovery and design relies on the use of pre-built datasets, therefore it is key that these datasets are properly assessed to ensure their reliability and quality.
Materials scientists have also begun to discuss the various ethical considerations that may come with incorporating AI into their work. This includes the importance of recognizing and mitigating any bias in an AI model’s training data, as well as remaining vigilant in recognizing AI-generated misinformation through the application of rigorous validation practices.