Machine learning has become a buzzword in all areas of science, and the hype around it can make it difficult to differentiate fact from (science) fiction. When it comes to researching new materials, some scientists believe we are now at the beginning of a second computational revolution at the hands of machine learning. But is that really true?

Materials science already went through a revolution in the 20^th century with the advent of computational methods such as density functional theory (DFT), Monte Carlo simulations, and molecular dynamics.

These methods have allowed researchers to simulate and study the properties of materials, which resulted in a significant reduction of the time and cost of developing new materials.  

There is now an explosion in the number of studies applying machine learning to materials science. Just five years ago, there were less than 200 scientific articles published each year on this topic — in 2023 alone, that number grew to over 2,000 papers published.  

So, what makes machine learning so interesting for materials research? The technology promises to overcome the limitations of the now ‘old-school’ computational methods, allowing scientists to run large-scale, complex simulations they could only dream of until now.  

When (and if) that happens, machine learning would be a game-changer for the development of better and more efficient materials for all sorts of applications including microchips, batteries, and solar cells among many others.

However, while machine learning is now commonplace in some areas of our day-to-day life — things like unlocking your phone with face recognition or doing a voice search — the technology is still taking its first (perhaps big) steps when it comes to materials research. This means there is still a lot of work to do until machine learning fulfills its potential. One day, machine learning might be an essential tool in materials science and lead us to incredible breakthroughs. But what can it actually do today?  

Studying a material’s crystal structure

A material’s crystal structure is the blueprint of how the atoms it is made of are arranged in a three-dimensional space. Identifying a crystal structure is a crucial milestone for understanding and designing new materials.  

However, studying a material’s crystal structure is much more complicated than you may think. Knowing which atoms make up the material is not enough. Why? There is a seemingly endless number of possible arrangements of the atoms within a 3D space.  

Computational methods that sort through all these possibilities require a significant amount of computational resources — which makes them costly and time-consuming, especially for complex systems.

Here’s where machine learning can make a difference.  

Machine learning algorithms can be trained with the data of known crystal structures to predict the crystal structure of a new material. Studies have shown that machine learning can achieve results as accurate as classic quantum mechanics calculations but with computational costs several orders of magnitude lower.  

This has the potential to significantly accelerate the study of crystal structures and take the discovery and design of new materials to a whole new level.

Predicting a material’s properties

The key to a material’s performance lies in its properties. When designing a material for a specific application, scientists will look for a series of properties, such as a whether the material is stable at a certain temperature range, how it interacts with light, how it responds to stress and strain, or how it conducts electricity.  

Machine learning algorithms can be used to predict how variations to a material’s composition or structure affect each of the desired properties, helping scientists find the best candidate materials for any given application.

For example, machine learning is already being used to design construction materials with better durability, sustainability and recyclability.  

Another promising area of materials research where machine learning is gaining traction is carbon capture. Here, machine learning can be used to predict the ability of a material to absorb CO2 under different conditions.  

Accelerating simulations

The most accurate simulations of a material are those that look at the behavior of individual atoms — these include well-established techniques such as DFT and molecular dynamics.  

However, there is a major bottleneck when using these techniques for high-throughput screening, which consists of scanning through thousands or even millions of material candidates looking for the ones that exhibit desired properties. The challenge is that running these simulations through millions of candidates takes enormous amounts of time and computing power.

Machine learning could help us overcome this limitation. By meshing machine learning algorithms with traditional simulation methods, their speed, and perhaps even their efficiency, can be significantly improved.  

This would unlock a whole new range of possibilities for scientists to study larger and more complex scenarios. Machine learning is already being used to improve high-throughput screening in the pharma industry to look for new drug candidates, and is now starting to do the same in materials research.

For example, this technology is having a big impact in the study of battery cells, which are complex systems where multiple materials interact with each other.  

The future of machine learning in materials science

Despite its promise to bring forth a second computational revolution to materials research, machine learning is still taking its first steps into this field. To fulfill its potential, there are some roadblocks we will have to overcome.

Another challenge will be translating academic research into real-life applications. While early machine learning prototypes have been able to identify millions of promising material candidates, there is no guarantee that those materials can be synthesized in the lab. And when it comes to industrial applications, some materials may require a production process too complex to be mass produced.  

Combining machine learning with robotics may be the way forward to overcome this challenge, allowing scientists to rapidly design, test, and optimize synthesis methods for new materials.

Moving forward, a major challenge is that training machine learning algorithms requires large amounts of high-quality, standardized data. One solution is for scientists to work together to create these datasets with experimental data, but that is a large and time-consuming task — and let’s not forget that experimental chemistry and materials science are often messy in real life. A promising alternative is to train machine learning models using simulated data. This is an approach that is gaining popularity in the field of artificial intelligence — and one of the approaches we follow here at Quantistry.

What we can expect in the longer term is for machine learning and computational tools to evolve together and merge, giving rise to more powerful, fast, and accurate models to study and design new materials. Scientists are already testing more advanced and powerful forms of machine learning in materials research, such as neural networks.  

At Quantistry, we work to bring the power of machine learning to the discovery and optimization of new materials. Our aim? To help promote sustainable futures across multiple fast-growing industries.  

We believe the revolution has already started, and our goal is to become a key part of it.

‍