Alán Aspuru-Guzik, University of Toronto
The Future of Materials Is Self-Driving
Written by Alison Hatt
In Monday’s plenary session, Alán Aspuru-Guzik gave an overview of his work on self-driving laboratories, in which he uses machine learning and robotics to accelerate materials discovery. He noted that 10 years ago at the MRS meeting, he and other researchers were focused on high-throughput virtual screening of materials. In a study from that time, Aspuru-Guzik used high-throughput screening to identify new materials for organic light-emitting diodes. His research team started with millions of candidates and gradually narrowed the list down to 1000 or so using machine learning and high-throughput calculations, but ultimately found that only 30 of them could be synthesized.
“It didn’t really move the needle,” said Aspuru-Guzik of the study. “We needed to actually accelerate discovery by not only calculating the materials but testing them.”
These efforts led to his vision for a self-driving lab. Instead of going serially through materials discovery process, he imagined a cycle of identifying candidate materials, synthesizing batches of them, and feeding the results back into the dataset continuously, so the machine learning tool can focus on synthesizable areas of parameter space.
But how to identify the candidate compounds? Using modern computational chemistry, we can determine the function of a material based on its structure, but a challenge for this century, according to Aspuru-Guzik, is to go the other way and determine the structure of a material to achieve a particular function.
In work published in 2018, Aspuru-Guzik and collaborators developed a tool for solving inverse design problems, using neural networks to encode a chemical structure, compress it into a bi-directional latent space, and then decompress it. In the latent space, the composition can be optimized by following gradients or doing sampling methods and, once a maximum is found, the material can be decoded to reveal the optimized structure. Many researchers are now using this generative model approach with various machine learning models, like genetic algorithms, generative adversarial networks, and, of course, large language models.
Demonstrating the power of the generative model approach, Aspuru-Guzik discussed a study in which he and collaborators developed a promising new drug candidate using artificial intelligence and generative models, going from ideation to testing in animals in just 45 days, usually a one-year process. He also shared examples from his work using a similar approach to design injectable materials for drug delivery and materials for carbon capture.
However, generative models can also dream up materials that are not synthesizable. With self-driving labs, Aspuru-Guzik hopes to better focus explorations on synthesizable materials space. As an example, he discussed efforts to discover new molecules for organic lasers, for which the screening criteria are a list of desired optical properties. Working with collaborators in British Columbia, Urbana Champaign, and Glasgow, his Toronto-based team took a modular approach, developing a library of molecular blocks and iteratively coupling them in varying configurations. The four collaborating teams all have self-driving labs, all of them slightly different, that can do synthesis, identification, and optical characterization of organic compounds. As a “hello world” experiment, the team produced 40 new organic laser compounds in just one weekend, several of them better than the molecule they started from.
More recently, the team demonstrated running the four geographically dispersed labs in parallel on a single discovery effort, sharing data and shipping materials between them. To achieve this, the researchers developed an asynchronous Bayesian optimization tool. They calculated 190,000 molecules with DFT, creating a model which they then calibrated experimentally by synthesizing and characterizing a hypercube sample of 500 compounds. Then, using their Bayesian optimization tool, the labs closed in on the most promising areas of materials space and ultimately identified 21 new compounds, now the brightest organic molecules in the world.
Looking forward, Aspuru-Guzik asked, what will the laboratory of the future look like? Anticipating a laboratory operated by independent robots, his group is training robotic arms to recognize transparent objects, which are particularly hard for robots to do but is an essential skill for working in a messy chemistry lab. They are also developing a computer program named Organa that integrates ChatGPT-4 technology and can plan and execute chemistry experiments given minimal input in natural language.
Aspuru-Guzik closed with a vision that computers can be agents of understanding, not just computational workhorses, a once-fantastical future that now seems almost within reach.
The Kavli Foundation is dedicated to advancing science for the benefit of humanity, promoting public understanding of scientific research and supporting scientists and their work.