How AI is becoming a research companion to materials scientists
By automating scientific processes and introducing artificial intelligence for decision-making, TRI’s new closed-loop research platforms free up scientists’ time for more creative tasks.
By - Shijing Sun
When I first started graduate school almost 10 years ago, I was mixing ingredients by hand, writing down reaction conditions on a piece of paper, and grabbing a quick lunch in between lab sessions. At that time, the idea of a robot doing my experiments — or using machine learning to predict the outcomes of my reactions — would have never occurred to me. I accepted a future as a scientist where I would only be able to explore a tiny fraction of the billions of possible materials in the universe by hand. If lucky, a scientific discovery might arrive serendipitously as I became better at making “educated guesses.”
My story did not follow this path. Owing to the recent, rapid development of robotic manipulation, computer processing power, and artificial intelligence (AI), researchers like myself are increasingly turning to data science and automation to augment their research. Within the past few years, I’ve performed high-throughput experiments at Massachusetts Institute of Technology (MIT) to screen promising solar-cell materials from thousands of candidates and applied machine learning to identify new materials within seconds. At TRI, I help develop AI frameworks that combine my materials and data science knowledge to guide scientific discovery. These experiences have been transformative for me and also are an illustrative example of the mindset shift for many scientists.
In this blog, I will discuss the Energy & Materials team’s recent efforts in building closed-loop platforms at TRI, which address key bottlenecks in moving towards fully autonomous research for materials science. Through our use cases in battery and fuel-cell materials, we observe that today, humans and AI are already working side-by-side to tackle some of the most urgent challenges of our times, rapidly making progress in materials development for a cleaner, more sustainable, and electrified future.
Teaching AI to make scientific decisions
Similar to how we take transportation every day to get to where we need to be, scientists drive their research vehicles — whether that is a piece of lab equipment or a computer — to reach scientific destinations. Having automated tools and processes makes a difference in speed, for example, a state-of-the-art research robot can work around the clock completing thousands of chemical synthesis without reducing the quality of work by getting tired, bored, or discouraged from early failures. However, even for a robot, it would be too expensive to simulate or synthesize every possible new material of interest. Therefore, it’s critical to save time and money by picking future experiments in an intelligent manner — a process referred to as optimal experiment design.
In the past, the scale and complexity of scientific experiments were limited to human brainpower. At TRI, we develop closed-loop platforms that leverage computer “brainpower” to achieve optimal experiment design. As recently described by the Energy & Materials team, by setting up AI frameworks that learn from past experiments and decide next steps, we teach computers to systematically make reliable scientific decisions without human intervention. As a result, we are not only automating the execution of experiments but also the whole scientific process, making materials research autonomous.
Closed-loop platform: a route planner for self-driving labs
Many materials science challenges today can be represented as optimization problems. TRI’s closed-loop platform solves these challenges by using computer science knowledge in the form of four steps: decision, experiment, hypothesis, and prediction (as shown in Figure 1). Once a set of new experimental parameters is determined in the decision step, automated data collection and analysis are initiated in the experiment step. The results are then fed into a machine learning model, outputting the system’s latest understanding of the material or device parameter landscape. Then, the predictive models are again combined with AI decision-making in the decision step to plan the next best “route” to take — in this case, a route might be the optimal fast-charging protocol for a long-lasting battery.
With this platform, TRI has demonstrated an accelerated materials discovery and design in three distinct areas: Computational Autonomy for Materials Discovery (CAMD), as described in the previous blog, focusing on the realization of novel inorganic crystalline materials; AI-Assisted Catalysis Experimentation (ACE), developing new catalysts for fuel cells; and Battery Evaluation and Early Prediction (BEEP), which is aimed at extending the lifetimes of batteries in electric vehicles.
Companionship: automation with a human touch
With fast-growing computation capacity, scientists continue exploring avenues using AI to better design new materials. To give an analogy to recent automated driving platforms, a fully autonomous material research platform would be like a vehicle aimed at full self-driving without human help (e.g. Toyota Chauffeur); whereas, the emerging physics/chemistry-informed closed-loop systems are more like semi-autonomous systems focusing on blending input from both humans and machines (e.g. Toyota Guardian).
One key reason for keeping humans in the loop is to allow for automated scientific processes to not only proceed quickly but also to provide scientific significance. For example, in our CAMD system, knowledge from machine learning, physics, and heuristics are seamlessly integrated into the hypothesis step. With this approach of automation with a human touch — a method called jidoka in Toyota Production System — we are creating new collaborations and work dynamics between AI and scientists on our journey to autonomous research.
You might be wondering: where does this new type of companionship in materials research drive us next? Stay tuned as we will be sharing some more exciting research from us at the crossroads of material science and AI in the near future!
