(ORDO NEWS) — A group of American researchers in the field of chemical engineering has developed a self-driving laboratory capable of identifying and optimizing complex multi-step reaction pathways for the synthesis of both new and already known materials and molecules.
During a concept demonstration, a neural network-driven system found a more efficient way to produce high-quality semiconductor nanocrystals that are used in optical and photonic devices.
Multistage synthesis of chemical compounds is a truly labor-intensive field of scientific research. It is not uncommon to develop a new target material or optimize a method for the synthesis of a particular chemical, requiring the work of dozens of specialists over several years.
At the same time, scientists are faced with the so-called curse of dimensionality: the more stages and reagents in the reaction, the exponentially more time it takes to enumerate all possible parameters of this reaction – combinations and ratios of volumes and concentrations of reagents, the time of their interaction, and so on.
Therefore, a completely logical direction of research was the use of machine learning methods with automated methods for setting up experiments in chemistry and materials science, which led to the creation of “self-driving labs” (self-driving labs, SDL).
Such systems, driven by neural network algorithms, are able to explore and solve problems in chemistry and materials science with incredible speed and efficiency. Neural networks are used here to correctly process the data of the previous experiment and select the optimal parameters for setting the next one.
Previously created SDL concepts, including entire laboratory rooms integrated with robotics and microfluidic reaction systems, remain highly specialized for solving problems with well-understood limited parameter spaces.
For SDLs to be truly mainstream, technologies must overcome two major barriers to dealing with complex, multi-step chemistry processes: the curse of dimensionality and lack of data.
A group of American researchers from the University of North Carolina and the University at Buffalo tried to solve these problems and developed AlphaFlow – SDL controlled by a neural network that learns from the reinforcement learning method.
In addition, AlphaFlow includes Modular Fluid Processing Units, high performance microfluidic flow reactors.
According to the authors, AlphaFlow is capable of autonomously and independently exploring, learning, and optimizing multistep reactions with parameter space complexity exceeding 40 dimensions, unlike previously used chemoinformatic and retrosynthetic experimental design methods.
As a demonstration of its capabilities, AlphaFlow has studied and optimized the reaction sequence for synthesizing cadmium selenide quantum dots with a cadmium sulfide shell without any prior preparation or knowledge of even the correct reagent addition sequence.
“We have shown that AlphaFlow can run more experiments than 100 chemists in the same time frame while using less than 0.01% of the relevant chemicals. It effectively miniaturizes and speeds up experiments, performing the same laboratory operations that would require an entire laboratory of experimental chemistry.
And all this – on a platform the size of a suitcase. Extremely effective,” said Milad Abolhasani, a professor of chemical and biomolecular engineering at North Carolina State University, the paper’s final author .
AlphaFlow is open source because scientists believe it’s important to share high-quality, reproducible, and standardized experimental data both successful and unsuccessful.
The researchers are now looking for partners in both the scientific community and the private sector to start using AlphaFlow to solve a wide range of chemical problems.
Contact us: [email protected]