Materials Nexus: Shaping the Future of Machine Learning for Net-Zero Materials

Humanity faces huge challenges with climate change and environmental pollution that won’t go away by simply installing a software update. Can computing still save the world?

One way forward may be to leverage computing to design better and more sustainable materials. Currently, material discovery involves a lot of trial and error in the lab – an arduous and manual process that takes many years and millions in funding and often yields only incremental advancements. 

Materials Nexus was founded by Jonathan Bean, Robert Forrest, Jon Pillow, and Nic Stirk in late 2020 to use computational modeling and machine learning to discover and optimize materials quickly and cost-efficiently. This addresses both enhancing existing materials, such as more sustainable magnets that go into electric motors, e.g., in wind turbines, as well as attempting genuine moonshots like discovering room-temperature superconductors. 

In the summer of 2023, Materials Nexus announced a £2M seed round led by Ada Ventures and joined by High-Tech Gründerfonds, The University of Cambridge, MD One Ventures, and several angel investors. 

Learn more about the future of machine learning for net-zero materials from our interview with the co-founder and CTO, Robert Forrest:

Why Did You Start Materials Nexus?

After hundreds of millions of years, Earth is built on a sensitive balance of different ecosystems to make it the habitable place we know. By burning fossil fuels at an unprecedented scale in our planet’s history, we are rapidly changing these ecosystems and pushing them out of balance. As dramatic as it sounds, as climate change progresses, humanity could well be pushed to the verge of extinction.

For years, I have put my research efforts into finding ways to solve this problem. Most other companies are either struggling to tackle it or simply ignoring it and are unwilling to take action; trading short-term economic gains for long-term potential catastrophe. 

Jonathan and I met at Cambridge. We were frustrated by the lack of impact from academic work and subsequent action in industry. We wanted to translate our work to “the real world” and make a genuine impact on the climate crisis. Through the Carbon13 Accelerator we met Jon and Nic, who shared our mission, and we founded Materials Nexus.

How Does Computational Modeling for Materials Work?

We’re part of a new wave of industry 4.0 and next-gen materials startups, leveraging big data and machine learning to solve materials science problems. 

It used to take decades and hundreds of millions of dollars to develop a new material, mainly because experimentalists had to spend a lot of time in the lab on trial and error. This is inefficient and not sustainable. We leverage machine learning to find and optimize materials before testing them, thereby minimizing the effort spent in the lab. 

One material example we are looking at is finding a more sustainable replacement for rare earth magnets. Rare earth elements are critical to magnetic materials and are typically imported from China, which has a de facto monopoly on its supply. This causes all kinds of geopolitical problems but is also problematic for the environment, as large amounts of energy are needed to extract and process the ores and transport them from the mines to the rest of the world, resulting in long supply chains.

If we can find a magnetic material that does not rely on rare earths and is as simple and easy to obtain as iron or nickel, it would fill an important gap in the market and help to avoid many environmental problems.

One of the main challenges for machine learning is gathering a lot of high-quality data. A starting point is the academic literature and publicly available materials databases, which help us understand what materials and properties are interesting. That way, we can bootstrap an initial machine learning model, benchmark its predictions, and get an overview of the problem space.

After identifying a material domain, we dive deeper. Literature data alone is then usually no longer of sufficient quality, so we generate our own data through computational modeling and quantum mechanical simulations. It is a continuous development process where we iteratively update the model as we gather more data and gain a better understanding of compositions and properties.

In the instance of magnets, we started with quantum mechanics calculations, i.e., solving Schrödinger’s equations to determine where electronic spins point in different materials, and then use machine learning to link elemental composition to differences in a material’s magnetic properties. 

For example, magneto-crystalline anisotropy can play a significant role in the performance of wind turbine motors. And soft magnets – materials that can easily magnetize and demagnetize – are another important category used, e.g., in speakers, transformers, and anywhere magnetic fields fluctuate.

We still use classical computers for the computational modeling and simulation part, as quantum computers are still too slow and error-prone to be viable right now. And even if their error rate gets reduced significantly, they will be limited in the size of systems – with machine learning, we can address systems comprising thousands of atoms.

Once we identify a promising magnetic material, we proceed to lab testing to verify its properties.

Ultimately, our benchmark is that it works – that we can design and synthesize an effective, new material. So it’s not just about composition but also about synthesis routes and processing steps – for example, if we develop a new alloy, we can also model the processing, annealing, and sintering steps. We are building a system that covers all length scales, from ab initio density functional theory simulations to technical length scales.

While we currently focus on magnetic materials, the longer-term aspirations for our system are far more ambitious. A prime example is to find a room-temperature superconducting material. There’s been a lot of hype this year around LK-99, which turned out not to be superconducting. With our system, we can attempt to discover new types of superconductors and maybe also ones that work at room temperature. 

How Did You Evaluate Your Startup Idea?

We started by looking at the current market for any given material, the current problems, the needs of customers, and its future growth.  For magnets, we know it’s currently a multi-billion dollar market and is poised to grow with the adoption of electric vehicles, electric motors, and batteries. We have now established a data-driven pipeline to consider all new material opportunities.

We will initially focus on a few very specific applications, such as magnets, before branching out further, e.g., into superconductors, catalysts, or corrosion-resistant materials. To address more material verticals, we’ll have to scale our technical team rapidly, particularly as we approach material development, a process we will be keeping in-house.

Creating a platform for developing new, superior materials and giving it away as a SaaS doesn’t really make sense – it’s like giving away the farm instead of the farm produce. Instead, we leverage it ourselves to build a sustainable business and focus on achieving our climate goals. That way, we can also ensure the system is used as intended and not to develop nefarious things. 

What Advice Would You Give Fellow Deep Tech Founders?

What is most important is to assemble a team that is much smarter than yourself and has a wealth of experience to innovate and drive the startup forward. People who challenge and continuously push to improve what you’re doing as a startup is critical. Building a startup is like cutting David out of the marble with your bare hands – it takes amazing grit and perseverance.