Teaching computers to “see” at the nanoscale

Assistant professor from UOIT and SOSCIP researcher is combining machine learning with computational modelling to accelerate the discovery of new materials important for renewable energy

Project Title: Machine Learning for Materials Discovery and Design
Researcher: Dr. Isaac Tamblyn, UOIT
Outcome: In 2017, the team will launch a start-up company to commercialize their results.
Supported by: SOSCIP, Federal Economic Development Agency for Southern Ontario, IBM Canada Ltd.

Advanced Manufacturing


From aerospace, to health, to energy, industries and researchers are racing to develop the next generation of advanced materials that can improve efficiency, enhance quality and extend the life of pre-existing materials.

Developing these ‘designer’ materials carries with it a heavy price tag, both in time and money spent. But Dr. Isaac Tamblyn, an assistant professor from the University of Ontario Institute for Technology’s (UOIT) Faculty of Science, is hoping to remove these barriers for future researchers by using recent advances in computer vision and artificial intelligence to reduce the cost and complexity associated with modelling them on a computer.

“Over the past decade, the field of computational material design has grown rapidly. Knowing only the elemental composition of a material, it’s now possible to simulate many of its properties (e.g. electrical conductivity, colour, how strong it is).”

“That said, if you want to trust the output of your computer prediction, it has to be accurate. Accuracy of course comes with a cost, and right now, even with huge computers, lots of interesting materials and problems are simply out of reach. Going forward, as engineers and material scientists seek to understand and control complex nanostructured materials, the problem is only going to get harder. If we really want to be able to do bottom-up design of a material, we need to be able to find a cheap and easy way to model them.”

Dr. Tamblyn is using SOSCIP’s state-of-the-art computing platforms, namely the Blue Gene/Q (BGQ) supercomputer and the Cloud Data Analytics platform to build an extensive material database for training advanced machine learning algorithms enabling accurate recognition of candidate materials in renewable energy applications. His research is supported by SOSCIP, a research and development consortium that pairs academic researchers and industry with advanced computing tools to fuel innovation leadership in Canada. With the support of SOSCIP’s high performance computing tools and expert personnel, Dr. Tamblyn is able to train a computer to make accurate approximations of difficult problems in material science.

Ultimately his group plans to use the technique to study new catalytic surfaces for use in clean energy applications, such as water-splitting and CO2 reforming.

The training approach is similar to Facebook and Google’s object recognition algorithms – where computers have been trained to recognize the faces of users in photographs by-way-of example.

“If you have lots of labelled photographs (e.g. cats, dogs, etc), with the right algorithm you can teach a computer to tell the difference between a cat and a dog. The process is very similar to how we teach children – you point at a cat and say ‘that’s a cat’, and then you point at a dog and say ‘that’s a dog’. Once a child has seen enough cats and dogs, they know the difference.”

“If you want to teach a computer the difference between cats and dogs, you need a lot of pictures of each. The same is true with materials. In order for us to develop approximate models of materials, we need to generate really accurate data first,” explained Tamblyn.

The research will provide a bright outlook for other researchers who are exploring how atoms behave together. “We’re coming up with a clever way of predicting properties so other researchers can deliver their results much faster,” he said, adding that currently, some types of calculations take as many as five million computing hours.

“If you have a BGQ lying around, that might not sound like a lot (it can do that in about 3 days), but not everyone has one in their backyard, so we’re trying to make this kind of work more accessible.”

At the moment, his team is investigating how to make carbon a better catalyst for water-splitting, a process that could be used to develop clean, renewable energy. Water splitting involves running an electric current through water to decompose it into oxygen and hydrogen gases, which can then be recombined to release energy. If successful, it could be used as a source of clean energy capable of powering electronic devices and fueling vehicles. Platinum is known to be a good catalyst for this reaction, yet it is not an earth-abundant material and therefore will never be viable at scale. Conversely, carbon is cheap and plentiful, but not efficient at the task. Carbon can take on many forms, however, so it’s possible that a modified form of it may hold the answer.

Dr. Tamblyn’s team includes a number of trainees, including SOSCIP-supported UOIT post-doctoral fellow Nataliya Portman. Dr. Portman’s role is to develop a combined mechanics and machine learning approach for prediction of material properties of molecular systems. She’s also gaining valuable skills which will make her uniquely qualified for a career as a data scientist.

“This project allows me to develop technical skills in high performance computing, GPU programming and big data management skills, and at the same time advance my knowledge and understanding of deep learning methods,” she said.

When the research is complete, Dr. Tamblyn is hoping to have perfected a validated machine learning model capable of rapidly determining the electronic structure and physical characteristics of catalytic surfaces. The team plans to use the model to launch a start-up company in 2017 targeted at industries that can benefit from high-tech material design, such as defence, aerospace and health.

“I think the biggest difference you can make is enabling lots of people to do great things,” explained Dr. Tamblyn. He describes the project as a sort of crowdsourcing of research resources. “I’m only able to move the ball forward so much. If you move it in the right direction though, and increase other researchers’ productivity [by building a tool that advances the research of others], you can have massive impact.”

SOSCIP is a research and development consortium that pairs academic and industry researchers with advanced computing tools to fuel Canadian innovation. SOSCIP supports projects that have the potential to have a considerable impact on the lives of Canadians, within areas such as water, cities, health and cybersecurity. The consortium includes 15 of Ontario’s most research-intensive academic institutions as well as Ontario Centres of Excellence and the IBM Canada Research and Development Centre.

fip_3lines_EN