TALK LIKE A MATERIALS SCIENTIST AND ENGINEER

Artificial intelligence (AI) — computer systems able to perform tasks normally requiring human intelligence

Compound — a mixture of two or more elements

Machine learning (ML) — computer systems able to learn and adapt without following specific instruction

Materials science and engineering (MS&E) — the study of materials’ compositions and properties, and developing materials to fulfil specific functions

Natural language processing (NLP) — computational techniques that can analyse and synthesise natural human language

Qualitative data — data that describe the attributes or properties of something

Quantitative data — data expressing a measurable property of something

Machine learning (ML) is a powerful emerging field involving programming computers that ‘learn’ as they go. A part of artificial intelligence, ML looks for patterns in existing datasets and uses the patterns it finds to make its own decisions or predictions. “Machine learning has undergone astonishing growth in its capabilities in the last couple of decades,” says Professor Dane Morgan, of the University of Wisconsin-Madison’s College of Engineering. “These new capabilities are impacting almost all aspects of life and the study of materials is no exception. Now, the challenge is for the field to figure out how to use it most effectively.”

Dane and his colleague Dr Ryan Jacobs are focusing on understanding ML’s capabilities and exploring how they can be harnessed within materials science and engineering (MS&E). MS&E involves the study of materials’ properties and the creation of new materials for specific purposes based on these relationships. “The use of ML in MS&E is in its early stages but evolving rapidly,” says Ryan. “We’ve seen its power in other fields – for instance, beating the world’s best professional chess or poker players, and enabling potentially transformative technologies such as autonomous vehicles – and it’s clear that ML is driving advancements in science and technology.”

A digital abundance

All areas of science have benefited from the rise of computation and the internet. “We now have easy access to papers and data, including massive digital databases that compile the structure and properties of countless compounds,” says Dane. “This is coupled with powerful computing and modelling capabilities that make it far easier to explore and predict complex relationships between materials’ structure, composition and their properties.”

The majority of this computing power comes from open-source software. “The fundamental ML algorithms are available to everyone for free,” says Dane. “This includes tools made by large corporations such as Facebook and Google, as well as packages made by other MS&E researchers.” This accessibility removes barriers and promotes exploration and development of further tools and discoveries.

Hidden correlations

ML is especially adept at finding patterns in huge datasets that would be impossible for humans to notice. “ML can identify incredibly complex relationships between materials’ structure, properties and performance,” says Ryan. “It extracts information from data to form links between a material’s composition (the types and quantities of elements it contains) and structure (positions of atoms in space) to particular properties or functions.” Dane gives the example of how the brittleness of steel, when exposed to certain conditions, is affected by the type and distribution of atoms within its structure. ML can be used to deeply assess these relationships, which is extremely useful for selecting a type of steel that will perform sufficiently in important roles – within a nuclear reactor, for instance.

As well as testing existing materials, these relationships can also be extended by using ML to predict structure-property-performance relationships in currently non-existent materials. “Molecular-scale simulations are routinely used to understand predicted properties prior to real-world experiments and are a key part of the discovery and design process,” says Ryan. “ML can learn how atoms in a material interact with one another based on existing data and uses this knowledge to calculate the properties of novel materials.” For example, ML is being used to discover which materials can be used to make solar panels as efficiently as possible, by predicting how structural changes to a solar cell absorber material affect its ability to convert sunlight into electricity.

Natural language processing

Though much existing MS&E data is quantitative – based on numbers that can be relatively easily compared – a lot of data is qualitative, which is more challenging to compare or even come across in the first place, given the vast number of existing MS&E research papers in existence. “ML is emerging as an effective means of data collection through natural language processing (NLP),” says Ryan. “Manually searching existing papers for relevant knowledge is massively time consuming, but NLP methods can extract information far more rapidly.” NLP involves using ML techniques to enable computers to learn the different ways that qualitative data might be presented in scientific papers – for instance, a preparation technique might say “the mixture is rapidly stirred” – and bring this data together into a format that can be used by scientists.

As such technologies become increasingly powerful, there is a growing need for MS&E researchers to adapt to use them effectively. “The new abilities of ML to understand language and make complex connections are still not widely used in MS&E,” says Dane. “There is a need to integrate the tools of ML into our existing workflows, to make our research faster, better and more cost-effective.”

Computer vision

Much of MS&E characterisation data is in the form of images. “In recent years, the scale and complexity of electron microscopy data has increased exponentially, driven in part by improved detector technology,” says Ryan. Researchers are often interested in determining which phases are present in a micrograph, or finding individual objects in an image, such as missing atoms or extended defects like dislocation loops or voids. Large, complex datasets necessitate the use of ML computer vision methods to automate this approach.

Computer vision ML algorithms encode the complex relationships present in images – like contrast changes or the presence of edges around an object – enabling the algorithm to ‘learn’ what a particular microstructure or defect may look like. “Once properly trained, we have found that computer vision methods can detect objects in images with about the same efficacy as human experts, but in a fraction of the time,” says Ryan.

The robot age

Like all scientific fields, MS&E inevitably includes tedious and time-consuming tasks, such as mixing hundreds of combinations of different elements to test their properties. Increasingly, robots are being developed that can perform these tasks, saving researchers a lot of time, effort and human and financial capital. “Entire autonomous laboratories are being developed that can make and characterise multitudes of materials with almost no human intervention, allowing rapid exploration of properties,” says Dane. “ML is excellent at performing the decision-making involved in this process, choosing which materials to explore and in what order, based on the data it receives as the experiments progress.”

Dane predicts the rise of such autonomous laboratories could fundamentally change the nature of MS&E. “As ML and robotics become cheaper and more effective, researchers will focus more on building and refining the autonomous environments, and less on the detailed experiments and calculations themselves,” he says. “This will change what skills MS&E researchers need.”

Ryan is enthusiastic about this changing nature of MS&E and science as a whole. “I am excited by the prospect that ML will become mature enough to be coupled with nearly every research task,” says Ryan. “I believe that each step of the scientific process, from hypothesis generation, extraction of existing knowledge from the literature, planning and running experiments and simulations, and analysing results will be aided by ML. This human-machine relationship will likely enable research to progress at a rate and precision that is presently unobtainable.”