Allegro-FM: AI Simulations Driving Sustainable Material Science

Artificial intelligence (AI) is making a lot of discoveries in the materials science field. All these different discoveries aren’t theoretical either, but real ones that are impacting a wide range of industries. 

For instance, Microsoft (MSFT ) utilized AI (Azure Quantum Elements tool) and high-performance computing to screen 32 million materials to narrow down¹ the candidates to just 18 in a matter of 80 hours, a process that would normally take several years. The scientists also synthesised a material that can reduce Lithium usage in batteries by 70%.

In another instance, MIT researchers used AI to design new proteins², which could be used to make materials with certain mechanical properties. These materials, which aren’t present in nature but are inspired by it, could potentially replace those made from petroleum and carry a significantly smaller carbon footprint.

Besides achieving breakthroughs in energy storage, battery innovation, and biomedical materials, AI is being actively explored by companies and scientists to advance semiconductor technology, aerospace, chemical manufacturing, and more.

Yet another critical usage of computational systems’ capability to perform tasks that are typically associated with human intelligence is being seen in the creation of sustainable and green materials.

An example of this is IBM’s (IBM ) RoboRXN, a project that combines AI, automation, and cloud to accelerate material discovery. With this remotely accessible chemical laboratory, IBM is enabling autonomous synthesis and testing of new molecules. The lab is being used to discover biodegradable polymers.

Researchers are also using machine learning to screen metal-organic frameworks (MOFs) for carbon capture.

Carbon capture is a key technology in reducing greenhouse gas (GHG) emissions from industrial facilities. MOFs here offer a promising candidate, thanks to their ability to selectively absorb carbon dioxide.

The molecules of these porous materials have organic nodes, inorganic nodes, and organic linkers, which can be arranged in different configurations, allowing for the design and testing of numerous potential MOF configurations.

To speed up the discovery process, researchers from the U.S. Department of Energy’s (DOE) Argonne National Laboratory used³ generative AI to quickly assemble over 120,000 new MOF candidates within 30 minutes. The calculations were run on a supercomputer with time-intensive molecular dynamics simulations conducted on the most promising candidates.

Now, researchers from the USC Viterbi School of Engineering are tackling the CO2 emissions of the concrete industry, which is responsible for about 8% of global carbon emissions, primarily through cement production.

The idea is to have concrete that not only self-heals and lasts even longer but can also trap carbon dioxide from the atmosphere; for this, the researchers have developed a revolutionary AI model.

This highly sophisticated model can simulate the behavior of billions of atoms at the same time, opening new possibilities for materials design and discovery at an unprecedented scale.

The Problem and the Solution: Carbon in Concrete

Every year, we see an increase in the frequency as well as the intensity of droughts, wildfires, rainstorms, and hurricanes. A major contributor to this is global warming, which is the result of increased greenhouse gas emissions.

GHG emissions refer to the release of gases like carbon dioxide (CO2), methane (CH4), and nitrous oxide (N2O) into the atmosphere. These gases trap heat and contribute to global warming and climate change. 

Now, the primary sources of greenhouse gas emissions include the energy sector, transportation, electricity and heat production, and agriculture. Manufacturing and construction are yet another major contributor, with building materials like concrete and steel having a large carbon footprint. 

A key element of modern infrastructure, concrete is everywhere around us, but the issue is that it presents a major problem due to massive GHG emissions.

Cement, which is a key component of concrete, is actually responsible for a substantial portion of the construction industry’s carbon footprint. The process of manufacturing cement involves the calcination of limestone, which releases large amounts of CO2, and the high-temperature heating of materials in kilns, requiring significant energy. 

Being the highest consumed product on earth after water, it is imperative to make concrete eco-friendly, or the environment will continue to be polluted with over a billion metric tons of CO2 annually due to this industry. A potential solution to this problem is trapping the harmful gas in the concrete itself.

“You can just put the CO2 inside the concrete, and then that makes a carbon-neutral concrete,” said Aiichiro Nakano, a USC Viterbi professor of computer science, physics and astronomy, and quantitative and computational biology.

So, Nakano, along with his colleagues at USC Viterbi, began researching “CO2 sequestration,” a process that captures and stores atmospheric carbon dioxide.

While a promising solution to achieve carbon neutrality by reducing GHG emissions, CO2 sequestration is a challenging process. To find the solution, the researchers turned to AI.

In their work, the team of researchers presented their foundation model (FM) for exascale molecular dynamics simulations. For this, they are taking advantage of E(3) equivariant network architecture (Allegro), a type of neural network, and a set of large-scale organic and inorganic materials data sets.

The resulting model is Allegro-FM, which can handle various material simulations for diverse downstream tasks.

How Foundation Models Are Transforming Materials Science

For their AI model, the team utilized the foundation model (FM), which it notes is a paradigm shift in AI and has transformed their way of model training. 

An FM is a large AI model trained on a massive, diverse dataset that can be adapted for different kinds of downstream tasks. These models are pre-trained, which means they do not require further training on a specific task. 

While generating an FM is extremely resource-intensive, one doesn’t have to start from scratch here; rather, the model can be adapted to an individual task through prompting or fine-tuning, thereby reducing both the cost and time required for model development and data generation.

A well-trained FM can be fine-tuned with much fewer resources, allowing one with modest computing resources to incorporate the model into their workflow.

As large deep learning neural networks, FMs are general-purpose models that can be used to build more specialized AI applications, including customer support, content generation, image creation and editing, language translation, healthcare, and robotics.

So, data scientists can simply use these models as a foundation to build their own machine learning (ML) models quickly and cost-effectively.

Microsoft-backed OpenAI’s GPT is one of the most common examples of FM. Its very first model was developed in 2018 using 117 million parameters, which was increased to 1.5 billion with GPT-2, released the next year. Then came GPT-3 in 2020, which has a 96-layer neural network and is trained using 175 billion parameters, which are adjustable numerical values including weights and biases that determine how models process input and output.

Now, its latest model, GPT-4 is rumored to have 1.76 trillion parameters, enabling generative AI apps to produce human-like content.

Other examples of foundation models include Amazon’s (AMZN ) Nova, Anthropic’s Claude, Google’s (GOOG ) Gemini, Meta’s (META ) LLaMA, and Stable Diffusion.

Given the many benefits of FM, the USC Viterbi researchers built the Allegro-FM, a foundation model for exascalable MD simulations. Their model is trained on publicly available large-scale Materials Project Trajectory (MPtrj) and SPICE datasets aligned with the Total Energy Alignment (TEA) framework for exascale materials applications.

The simulation of materials tends to require a large number of atoms to describe key features like distinct phases, phase boundaries, and grain boundaries, as well as an accurate description of chemical reactions. 

Allegro, according to the study, has great potential and is particularly suitable for materials simulations that require millions to billions of atoms.

Allegro-FM’s Breakthrough: Simulating Billions of Atoms

Published in The Journal of Physical Chemistry Letters, the study details⁴ the AI model Allegro-FM and its superior capabilities. The AI-driven simulation model comes with great capability to simulate billions of atoms at once.

By testing different concrete chemistries virtually before testing them in the expensive real-world experiments, Allegro-FM can advance the development of concrete, not as a carbon source, but as a carbon sink.

The models’ scalability is of key significance here. The thing is, many molecular simulation methods already exist, but they are limited to just millions, if not thousands, of atoms.

In contrast, Allegro-FM can handle more than four billion atoms while showcasing a 97.5% efficiency when simulated on the Aurora, an exascale supercomputer that can perform over a quintillion calculations per second.

This is a huge feat, representing computing capability about 1,000x larger than what older models can handle.

Moreover, the new model is flexible, covering 89 chemical elements. So, it can be used to predict molecular behavior for a vast variety of applications, such as cement chemistry and carbon storage.

“Concrete is also a very complex material. It consists of many elements and different phases and interfaces. So, traditionally, we didn’t have a way to simulate phenomena involving concrete material. But now we can use this Allegro-FM to simulate mechanical properties (and) structural properties.”

– Ken-Ichi Nomura, a USC Viterbi professor of chemical engineering and materials science practice, with whom Nakano has been collaborating for over two decades.

The simulations run by the researchers show Allegro-FM to be carbon neutral, making it a more suitable option than other concrete. Interestingly, the AI model solves yet other problems with concrete. 

The fire-resistant material, which is a major emitter of CO2, only lasts about a century on average. Unlike modern concrete, ancient Roman concrete has lasted far longer, for over 2,000 years. 

The recapture of CO2 can also help make the concrete more durable.

“If you put in the CO2, the so-called ‘carbonate layer,’ it becomes more robust.”

– Nakano 

So, Allegro-FM can simulate a carbon-neutral concrete that can last much longer than its current lifespan.

AI-Powered Cement Design with Allegro-FM

For the experiment, the team applied Allegro-FM to a tobermorite 11Å (T11A) crystal, a calcium silicate hydrate (CSH) mineral commonly found in nature. This particular mineral is found in ancient Roman concrete.

Cement is gaining a lot of attention as a carbon-storage material due to its ability to trap carbon through the mineralization process. CO2 mineralization occurs naturally as part of the geochemical cycle. A mechanistic understanding of this is still crucial to ensure long-term and safe carbon storage. 

Here, advanced simulations can provide insights, but they are either computationally expensive or limited in chemical accuracy. Scalable foundation models, however, show promise for accurately modeling cement’s carbon storage behavior.

So, Allegro-FM, equipped with the generalizability, scalability, predictability, and computational efficiency, shows significant promise in enabling dynamical simulations without sacrificing the details at the atomic scale.

According to the study, the framework can be utilized in the investigation of the nanostructure of CSH gel, self-healing cement, and durable cement design, providing a novel approach for geophysical science and civil engineering applications. It stated:

“Allegro-FM exhibits excellent agreement with high-level quantum chemistry theories in describing structural, mechanical, and thermodynamic properties, while exhibiting emergent capabilities for structural correlations, reaction kinetics, mechanical strengths, fracture, and solid/liquid dissolution, for which the model has not been trained.” 

The Future of Materials Research: AI Replacing Quantum Mechanics

The usage of AI has completely changed the game for scientists and engineers. Typically, simulating the behavior of atoms would have required a precise series of mathematical formulas or “profound, deep quantum mechanics phenomena.” And that was not only very technical but also slow.

But not anymore. The advent of AI, more specifically the advancement in the technology over the last couple of years, has transformed how scientists conduct their research.

“Now, because of this machine-learning AI breakthrough, instead of deriving all these quantum mechanics from scratch, researchers are taking (the) approach of generating a training set and then letting the machine learning model run.”

– Nomura 

The models can now handle vast volumes of data as well as more complexity while using fewer resources. This has made the entire process much quicker and efficient. 

In fact, Allegro-FM can predict “interaction functions” between atoms accurately. How atoms react and interact with each other normally requires lots of individual simulations and countless hours of calculation.

But instead of doing one element at a time, with each one having different equations and several unique functions, the new AI allows for the simulation of interaction functions with the entire periodic table at the same time, which means faster simulations and more material options. Nomura explained:

“The traditional approach is to simulate a certain set of materials. So, you can simulate, let’s say, silica glass, but you cannot simulate (that) with, let’s say, a drug molecule.” 

Technologically, as well, the new system is far more efficient with AI models able to make precise calculations that used to require a large supercomputer. This simplifies tasks and allows supercomputers to be used for more advanced research.

As Nakano noted, the AI can “achieve quantum mechanical accuracy with much, much smaller computing resources.”

This study shows some brilliant results, but there’s still a lot more work to be done. As Nomura said, the team will continue their concrete study research, moving to “more complex geometries and surfaces.”

Investing in Green Chemistry

One of the largest U.S. chemical companies, Dow (DOW ), actively invests in sustainable materials and carbon capture technologies. AI-driven simulation like Allegro-FM presents a perfect tool for companies like Dow to develop greener chemicals. 

Dow Inc. (DOW )

The company operates through a few different segments, including Packaging & Specialty Plastics, Performance Materials & Coatings, and Industrial Intermediates & Infrastructure.

With a market cap of $17.72 billion, the shares of Dow Inc. are trading at $25, down 37.53% YTD. It has an EPS (TTM) of 0.40 and a P/E (TTM) of 62.66. The dividend yield available is 5.58%.

This week, it reported its Q2 2025 financial results, which revealed a 7% YoY decline in net sales to $10.1 billion and 1% YoY drop in volume. 

The company’s GAAP net loss came in at $801 million and GAAP loss per share of $1.18. Cash from operating activities was negative $470 million. During this period, Dow returned $496 million to its shareholders through dividends.

This year, the company partnered with Google to develop an AI sorting system to identify film plastics and mixed materials. The system will utilize Google’s ML models, Dow’s material design knowledge, and recycling technology from Dow’s recycling company Circulus, in order to identify recyclable materials based on their molecular composition.

Dow also utilizes AI to optimize its chemicals and materials manufacturing processes. The technology enables it to monitor and adjust variables in real-time, ensuring efficient processes and consistent results.

Moreover, it employs AI in its R&D efforts with an aim to not only discover new materials but also improve its existing ones.

Latest Dow Inc. (DOW) Stock News and Developments

Conclusion

Concrete is a powerful material, utilized extensively in construction due to its strength, versatility, cost-effectiveness, and durability. But it is also responsible for emitting a significant amount of carbon dioxide, making it critical to look for alternatives. 

The latest study with its Allegro-FM tool has achieved a massive breakthrough in scalability for materials research that can enable simulations of billions of atoms at once, allowing us to rethink materials we depend on every day, like concrete. 

By offering a new way to design cleaner, longer-lasting materials like concrete, AI can help create not just stronger buildings but also a cleaner environment, leading to a future where materials like concrete aren’t part of the climate problem but rather a solution!

Click here to learn how we can overcome concrete’s shortcomings. 

References:

1. Chen, C.; Nguyen, D. T.; Lee, S. J.; Baker, N. A.; Karakoti, A. S.; Lauw, L.; Owen, C.; Mueller, K. T.; Bilodeau, B. A.; Murugesan, V.; Troyer, M.; Thien, D.; (…additional authors as listed). Accelerating computational materials discovery with artificial intelligence and cloud high‑performance computing: from large‑scale screening to experimental validation. arXiv preprint arXiv:2401.04070 (2024). https://doi.org/10.48550/arXiv.2401.04070
2. Anand, N.; Eguchi, R. R.; Mathews, I. I.; Höhn, M.; Chen, Y.; Pellegrino, J.; Li, J.; Lee, Y.; Park, J.; Zhang, J.; Baek, M.; Ovchinnikov, S.; Baker, D. Protein structure and sequence design with equivariant denoising diffusion probabilistic models. Chem, 9(6), 1646–1664 (2023). Published online April 20, 2023. https://doi.org/10.1016/j.chempr.2023.03.013
3. Park, H.; Yan, X.; Zhu, R.; Huerta, E. A.; Chaudhuri, S.; Cooper, D.; Foster, I.; Tajkhorshid, E. A generative artificial intelligence framework based on a molecular diffusion model for the design of metal‑organic frameworks for carbon capture. Communications Chemistry, 7(1), Article 21 (2024). Published online February 14, 2024. https://doi.org/10.1038/s42004‑023‑01090‑2
4. Nomura, K.; Hattori, S.; Ohmura, S.; Kanemasu, I.; Shimamura, K.; Dasgupta, N.; Nakano, A.; Kalia, R. K.; Vashishta, P.
Allegro‑FM: Toward an Equivariant Foundation Model for Exascale Molecular Dynamics Simulations.
5. The Journal of Physical Chemistry Letters, 16 (25), 6637–? (2025). Published online June 20, 2025. https://doi.org/10.1021/acs.jpclett.5c00605