Orbital Materials: Shaping the Future of Machine Learning for Green Material Design
Large language models have amazed us with their linguistic capabilities, stringing together words to compose poetry, marketing copy, and business emails. However, machine learning has even more potential to address pressing global challenges.
Imagine a world where aviation relies on sustainable fuels, batteries power vehicles more efficiently and sustainably, and carbon capture technologies aid in combating climate change. This can become a reality by using advanced machine learning models to better understand the world of atoms and develop better materials.
Orbital Materials is at the forefront of this movement, developing a foundation machine learning model to design green materials. Founded by Jonathan Godwin, Daniel Miodovnik, and James Gin-Pollock in 2022, it raised a $4.8M seed round from Fly Ventures, Compound, Flying Fish Partners, Character, and Mustard Seed Maze.
Learn more about the future of machine learning for green material design from our interview with the co-founder and CEO, Jonathan Godwin:
Why Did You Start Orbital Materials?
In my previous work at DeepMind, I witnessed the extraordinary ability of machine learning models to solve important physics problems. AlphaFold solved the protein folding problem, which will lead to the development of better drugs and provide cures for many previously incurable diseases. That got me excited: maybe we could use machine learning models also to solve the most pressing problems related to climate change and sustainability.
In particular, using machine learning to develop better materials seemed to be neglected. Some academic research groups and corporate innovation departments were looking into it, but they weren’t moving as fast as they needed to in order to make a dent. Few others seemed to be taking a startup approach, but I felt that the startup culture offered great potential for connecting the two different worlds of machine learning and materials science.
It’s an extremely rapidly evolving field that went from everyone assuming, ‘this doesn’t work,’ to solving material science problems within a few months that people have worked on for decades. Startups were the right kind of company to match that rapid development pace. From the moment AlphaFold and GPT-3 came out, I thought about creating a startup myself—and so I quit my job at DeepMind, and here I am, building Orbital Materials.
How Do You Leverage Foundation Models For Material Design?
Imagine you want to capture a pollutant from water. You need to design a suitable filter, i.e., at the atomic level, you need to find an arrangement of atoms that looks like a cage, lets the water through, and only traps the pollutant. This is an interesting but difficult chemical challenge since you only want to trap the pollutant. There are millions of possible materials you might try—and most of them won’t work.
To avoid spending a lot of resources on prototyping, you might try to model the problem on the computer with a simulation. However, if you start bottom-up with the physical equations, you soon get stuck because the dynamics are just too complicated, and it requires an enormous amount of compute to solve the equations numerically.
What we learned from AlphaFold is that machine learning can be incredibly useful in solving problems that arise from the interaction of many individual components to produce novel, emergent behavior. For example, the interactions between the individual atoms in a protein determine how the protein folds, resulting in its overall shape. Similarly, the atoms in such a cage would have to fit together to capture only the pollutant.
Instead of solving physics equations numerically, you replace those simulations with the prediction of a generative AI model. The core idea of generative AI is that you give it a specification of what you want and then let it generate examples according to that specification. So first, you feed the AI model the specification by training it on previous examples, and over time it learns to find new examples on its own. We will train a model to generate material for applications it has never seen before, such as capturing CO2 from the atmosphere, or removing a pollutant from water.
When developing a machine learning model, there are two main things you need to decide: The type of model—e.g., language or image models—and the architecture of the model. For example, ChatGPT is a large language model and uses the transformer architecture that Google described in a publication in 2017. Stable Diffusion is a diffusion model and uses a convolutional neural network architecture. At Orbital Materials, we are adapting these models to make them more physically informed so they can be used for materials science and chemistry.
Developing the machine learning model is only part of our overall work. Everything we develop has to be useful to later go into a lab and produce an actual material, so we spend a lot of time designing high-quality benchmarks and tests based on experimental results. Some benchmarks already exist, but we also put a lot of resources into this work ourselves.
Our ultimate goal is to develop materials that are better than the current alternatives. And clearly, the customer doesn’t care what method we use to achieve that goal, whether it’s machine learning, quantum computing, or something else. Nevertheless, there are good reasons why we may stay with machine learning in the long run:
Firstly, obtaining a simulation result via machine learning is incredibly fast; it takes milliseconds compared to days for normal physics simulations. So far, I haven’t seen quantum chemistry simulations run that fast on a quantum computer or simulate tens of millions of atoms. Yes, you could get better accuracy on a quantum computer for ground state energy, but for a large fraction of materials improvements in the ground state accuracy are not that relevant.
If you think about highly entropy alloys, for example, it’s more about getting the distribution of atoms right and not about determining the energy level of a specific configuration with high accuracy. So I think people are fundamentally overestimating how much of an impact improving the electron volt accuracy of electronic structure calculations will have.
Second, it’s hard to move materials from the lab to the application. How do you optimize automated process design? This is a combinatorial problem with a fair amount of data, and machine learning is really good at solving such problems. Wherever data live, machine learning lives. It’s useful not only for suggesting new material candidates but also for optimizing process parameters and costs for new manufacturing routes. So, it’s a lot broader in its application set than quantum computing.
How Did You Evaluate Your Startup Idea?
The 20th century was shaped by humanity’s success at solving material science problems: Think of building microchips and developing suitable lithography methods or being able to feed the world as we developed ammonia fertilizer.
When you think about where we want to be as humanity at the end of this century, it’s about doing all the things we like to do without destroying our planet: Driving cars that don’t pollute. Drinking clean water. Flying to Mars in a spaceship. We want to be living in a world without fossil fuels, where we can go to the stars, with non-invasive medical devices and revolutionary computer chips, and for that, we will need to solve difficult material science problems.
So there was never any doubt that there was a business. Rather, it’s about how to execute on that business opportunity, capture value, and build a large business to have a large-scale impact. That’s the challenge: to start small and grow as you provide and capture value along the way. We paid a lot of attention to what others have failed at in the past and developed our own unique approach to business building.
What Advice Would You Give Fellow Deep Tech Founders?
It’s easy to be intimidated by how difficult it can be to build a business and solve all these challenges. However, we consider it a privilege to have investors funding us to work on these challenges. Don’t think too much about how hard it will be, but recognize it as a privilege to build a startup. And for that, work on a problem that you truly enjoy solving.