MIT Professor of Mechanical Engineering Tonio Buonassisi says with AI, businesses don’t have to keep using trial and error for systems that take several months to build. Instead, machine learning can reproduce the sort of advances that is being made at MIT and SMART in the comfort of your own office, and, can help find a solution exponentially faster than one may have done otherwise…
By Tonio Buonassisi
Imagine you are in charge of research at a large multinational corporation and your boss presents you with a challenge: Your mission is to develop a new, sustainable material to enter your supply chain that must meet specified targets of cost, manufacturability, reliability, sustainability and performance. And because the climate crisis we face is urgent, you need to figure this out within 2–3 years and on a tight budget.
Scientists and engineers face challenges like this almost daily, in an effort to make renewable energy, biodegradable plastics, and disease-fighting molecules to make human society more sustainable. This is challenging because a vast number of materials need to be tested before you can find the right one that achieves all requirements. Most combinations will be bad, only a few will be good, and your job is to find a needle in the haystack that will satisfy your KPIs with your boss — and help save the planet. Because the 81 non-radioactive elements on the periodic table combine in trillions of unique ways, “brute-forcing” materials discovery can take decades, if not longer.
Since 2013, the research being done by my team at the Massachusetts Institute of Technology (MIT) and SMART, the Singapore-MIT Alliance for Research and Technology, MIT’s research enterprise in Singapore, and our research collaborators, has been using artificial intelligence to do just this, namely developing new materials for applications to benefit society.
What has emerged through our research is a combination of automation, data science, and computation — building blocks that promise to accelerate the rate of new materials development tenfold, and eventually millions of times, faster.
Indeed, less than a decade since we began, we have brought together these building blocks to deliver a blistering pace of materials development, which can be applied to some of the most pressing challenges of our time.
The Power of Automation
Automation is the first building block in our approach. Automation allows us to conduct experiments faster, speeding up the pace of research. Taking the example of assessing combinations of materials, where we started at a pace of 2 samples per week in 2013, today, we are looking at 336 samples per week — over a hundredfold acceleration.
Automation is so powerful for scientific research, because many materials, like the crystals inside of solar panels, cell phones, and LED lights, are built of repeating motifs, like LEGO blocks that can be assembled in different ways. Automation accelerates the building, screening, and testing of different material motifs. We use automation to prepare a variety of precursors, which can then be tested using standard experimental protocols to assemble those material motifs in a more organised fashion, and eventually test them to see which combinations meet required specifications.
In this way, the number of unique materials that we are making is rather stunning. In a period of around two months, we discovered recipes to make four new thin-film materials, and two new classes of semiconducting materials that hadn’t been reported previously in literature — something that probably would have taken us a few years to do previously.
An additional benefit of automation is greater reproducibility. Unlike humans, robots don’t get tired or sloppy. Tighter process control enables a better signal to noise, which means fewer experiments are needed to test a hypothesis.
Computation: Focusing Our Effort For Greater Impact
The second building block is high-performance computing. Moore’s Law affords us ever more powerful computers, which are getting better at simulating the real-life behaviour of materials. We use these computers to narrow down the number of possible candidate materials, somewhat analogous to how one might use a Google search to narrow the number of candidate websites when searching for some information.
Instead of testing every available material, we can focus our experimental bandwidth on just the ones that our computational models suggest are the most promising. In the best case, this decreases the number of experiments we need to run from trillions to thousands.
Computation can also help us better “represent” a new material in a machine learning algorithm. It’s not trivial to take a 3-dimensional antibiotic molecule, for example, and communicate its essence to a machine learning algorithm. Efficient computation is helping us to make better antibiotics by helping us discover more meaningful ways to represent their essence in machine learning algorithms.
Machine Learning: Helping Humans Make Sense of Complex Data
Machine learning is the third building block. The amount of data that high-throughput experiments and computation can generate is truly overwhelming, up to 20 gigabytes per hour.
We use machine learning to help make sense of this data, and decide what experiments to run next, in a statistically rigorous way. We use machine learning to decide what materials to make, how to construct them, and when to test their performance, cycle after cycle.
To highlight how incredible this closed-loop experimentation is, out of approximately 5,000 different combinations of materials to test, our algorithm has managed to converge on the most stable and best-performing materials in just under 100 attempts, and that is just from scanning through a fraction of the parameters available. This resulted in a new perovskite solar cell material that is more durable than state-of-the-art perovskite solar cell materials.
If you have bad data feeding into your model, your machine learning model predictions won’t match reality. That’s why we combine both experiment, theory, and human domain expertise into our models.
Every experimental cycle, the previous rounds of experimental data are fed into the machine learning model, in addition to thermodynamic simulations. This has to balance the experiment, since there can be a gap between simulation and the real-world findings.
The machine learning model has to learn how to take different streams of information and blend them together to make a decision about what samples to try next. The good news is that such forms of data fusion actually work, and help us learn faster.
This act of blending experiments and scientific theory helps us learn faster and ensures our findings are grounded in scientific theory; this is alternatively called “science-aware machine learning” or “physics-informed machine learning.”
AI for a better world
Yet it is not enough to have a new material in the laboratory and even make a device out of it; at the end of the day, our work centres on the world we live in and how we can have a positive impact on human life.
For instance, I believe that every human being on the planet should have access to clean drinking water to lead a healthy and productive life.
One of the things we have been doing, in collaboration with other researchers in the mechanical engineering department at MIT, is developing solar-powered desalination units that take brackish water and convert it into drinking water.
One of my team members interviewed more than 50 chief technology officers and leaders of research and development at different startups and multinational companies to understand the factors that limit system development the most. She found that the prototyping stage is the single point of development that takes the longest. Because of this, we have been developing technologies to accelerate prototypes so you can make them 100 times faster.
In the case of the solar-powered desalination units, these devices comprise solar panels, batteries and membranes that separate the salts from the water. In the end, drinking water comes out of these units.
That process can be represented mathematically and inserted into an optimisation algorithm in a machine. By doing this, we are able to drive down the time and cost of developing a prototype by optimising the sizes and specs of the solar panels, membranes, batteries and all the other components within the computer model.
This means we don’t have to keep using trial and error for systems that take several months to build until we realise we got it all wrong and have to start again. We can now do that whole design phase on a computer. This process is called generative design.
Through our work, we are also trying to impact the lives of many more people through education. We have designed an online series of lectures that take our code and datasets and offer them so you can run them on your computer and come up with a solution to your boss’s challenge of finding a viable sustainable material for your business.
You can reproduce the sort of advances that we have been making at MIT and SMART in the comfort of your own office, and then bring them into your research and development to find a solution exponentially faster than you might have done otherwise. We’re working smarter than ever before, to solve humanity’s biggest challenges, and we invite you to join us.
(Ed. Prof. Tonio Buonassisi is a Principal Investigator at the Low Energy Electronic Systems interdisciplinary research group at Singapore-MIT Alliance for Research and Technology (SMART) and a Professor of Mechanical Engineering at MIT. Buonassisi founded the Accelerated Materials Development Programme at Singapore Agency for Science, Technology and Research and founded the MIT Photovoltaics Research Laboratory. His work was recognized with a Presidential Early Career Award for Scientists and Engineers, among others. Featured image of Dr. Shijing Sun and Buonassisi discuss thin-film solar cell manufacturing in the MIT lab, courtesy of John Freidah, MIT Department of Mechanical Engineering.)