If you have an ACS member number, please enter it here so we can link this account to your membership. (optional)

ACS values your privacy. By submitting your information, you are gaining access to C&EN and subscribing to our weekly newsletter. We use the information you provide to make your reading experience better, and we will never sell your data to third party members.



This self-driving lab learns to make nanoparticles all on its own

The machine learning–guided platform called AlphaFlow discovered and optimized a new synthesis for quantum dots

by Ariana Remmel
April 17, 2023


A diagram in three parts going left to right. The left image is a schematic of a microfluidics reactor. The middle image is a computer with two arrows cycling around it, and the label says "reinforcement learning algorithm." The right image is a branching tree of nodes with a highlighted path from the very top node to a spot at the bottom, and the label says "Multi-step synthesis."
Credit: Will Ludwig/C&EN
AlphaFlow's microfluidcs reactor trains a reinforcement learning algorithm to discover and optimize the multistep synthesis of nanoparticles without human intervention.

Designing new reaction sequences to build complex materials can be a slow process, especially when each step involves multiple variables. A new machine learning–guided microfluidics platform could speed up that process by discovering and optimizing its own multistep syntheses (Nat. Commun. 2023, DOI: 10.1038/s41467-023-37139-y).

Researchers have previously combined artificial intelligence and automation to build self-driving scientific instruments to navigate the expansive reaction space. But it’s been difficult to compile sufficiently large and high-quality data from the literature to train the algorithms that drive these machines, which still require some level of human intervention to perform successful syntheses. To address this challenge, Milad Abolhasani, a chemical engineer at North Carolina State University, and his colleagues set out to create a miniature, self-driving microfluidics platform for multistep syntheses that could generate its own data and learn from its results.

So Abolhasani and his team built a briefcase-sized device called AlphaFlow. This device performs reactions in 10 µL droplets. It continuously monitors the microdroplets using spectral analysis techniques such as photoluminescence and absorption measurements to evaluate the success of each reaction. AlphaFlow then devises sequences of reactions by trial and error to create its own map of the potential reaction space. “It enables us to increase our data generating throughput by orders of magnitude compared to a conventional way of performing reactions,” Abolhasani says. AlphaFlow uses this data to train a machine learning algorithm via a technique called reinforcement learning. This approach allows the device to predict not only the immediate outcome of each reaction step but also how an individual step could impact the synthesis later in the sequence.

To demonstrate AlphaFlow’s self-driving capabilities, the research trained the device to synthesize cadmium-based quantum dots, which are a type of semiconductive nanoparticle, using a process called colloidal atomic layer deposition (cALD). This complex fabrication method works by iteratively depositing individual layers of specific semiconductor materials onto a spherical core. Abolhasani and his team gave AlphaFlow the spectral measurements of the cadmium-based quantum dots, the precursors needed to perform the synthesis, and control over more than 40 variables to adjust the reaction conditions—giving the device a discovery space of more than 1012 possible sequences to explore. But the researchers did not give the device any clues about how humans make the desired product, Abolhasani says. Instead, the team directed AlphaFlow to figure out how to make the quantum dots on its own and then apply what it found to devise the most efficient synthesis.

After around 30 days with zero human intervention, AlphaFlow produced the desired cadmium-based quantum dots via a novel recipe that is two steps shorter than conventional methods and eliminates a washing step that researchers previously considered crucial to the process. “Using reinforcement learning, the algorithm we use for AlphaFlow could look at the delayed rewards of its actions, and that’s why it could actually discover a new chemistry for [cALD] that was better and much more efficient than human-discovered chemistry,” Abolhasani says.

“AlphaFlow represents a significant breakthrough in automating complex chemical reactions,” writes Suhas Mahesh, a materials physicist at the University of Toronto, in an emailed statement. “The surprising aspect of the paper is AlphaFlow’s ability to choose steps that might not seem beneficial initially but lead to better outcomes in the long run,” writes Mahesh, who was not involved in the study. “This work is a very nice example of allowing a platform to choose between an intractably-large number of possible trajectories, in this case defined by a multistep cALD process,” Connor Coley writes in an email. Coley, a computational chemist at the Massachusetts Institute of Technology who was not involved in the study, says that AlphaFlow “takes the field one more step toward truly autonomous laboratories.”


This article has been sent to the following recipient:

Chemistry matters. Join us to get the news you need.