With arms, fingers and a machine-learning-powered brain, cobots make the humdrum routine work in chemical laboratories efficient — and exciting.
The roots of artificial intelligence have sprawled and gripped many facets of daily life — everything from the software powering our smartphones to self-driving vehicles and the various work applications that unleash a new wave of productivity.
These AI roots have now wriggled their way into the scientific research laboratory. Imagine slender robotic arms mixing chemical concoctions without spilling one single drop, transferring them deftly from one piece of equipment to another, while an analyser, infused with a machine-learning (ML) algorithm, patiently awaits samples for performance testing — all masterminded by artificial neurons firing away in a desktop sitting in a researcher’s office.
Associate Professor Daria Andreeva, a principal investigator at NUS I-FIM, is bringing this visionary concept to life. In a new paper published in the journal Advanced Intelligent Systems, Assoc Prof Andreeva’s research team, with collaborators from ITMO University, demonstrated that collaborative robots, or “cobots”, can be programmed and automated to synthesise graphene oxide (GO)-based membranes of various configurations, with their properties ascertained and predicted using ML-based algorithms. The membranes were previously developed by NUS I-FIM researchers, with details published in the journal Nature Nanotechnology.
While prototyped to produce membranes, the team’s setup can potentially be extended to “robotise” the preparation of other chemicals. This could help address various bottlenecks in basic research — particularly those involving laborious laboratory work — making the discovery of new compounds, or finding better ways of making old ones, many times quicker.
A robotic sleight of hand
“Cobots are equipped with torque sensors and power limitations that make them safe to work near humans while handling expensive, fragile lab equipment,” said Assoc Prof Andreeva. “For lab researchers, they’re the ideal ‘partner-in-crime’.”
This is because repetitive, mundane tasks can be delegated to cobots, who possess far greater physical and mental endurance than humans. They never — or cannot — complain, too. This liberates much of a researcher’s mental capacity, shifting focus to other tasks that require higher cognitive skills.
While robots have been resident in laboratories for decades — think electronic multichannel pipettes or automated microplate readers — cobots that could perform entire experiments, from synthesis to processing to analysis, is a different playing field.
To demonstrate this, Anna Nikitina, an intern in the Assoc Prof Andreeva’s group enlisted a cobot to produce graphene oxide-based membranes. Membranes are inseparable from many modern processes, whether filtering harmful substances from the air or recovering valuable resources from waste. Devising more efficient methods to produce them is, therefore, always welcome.
Operating autonomously for over 24 hours, the team’s two-armed cobot, animated by robotic programming, managed everything from the initial mixing of reagents to the careful handling and processing of the membranes. The setup can be segmented into five principal ‘blocks’: reagent mixing, centrifugation, vacuum filtration, drying and analysis. For analysis, the membranes underwent a permeability test of lithium, sodium, potassium and caesium cations by varying the pH of the permeate solutions. Witness the cobot in action by scanning a QR code in the researchers’ paper (Figure 1).
The beauty of this approach lies in its modularity — each ‘block’ can be independently adjusted and controlled. This not only enhances the adaptability of the process but also allows the cobot to tailor the properties of membranes to meet specific needs. A change in the pH or the chemical makeup of the membrane can dictate which cations are allowed to pass through.
“Furthermore, the consistency provided by robotic automation lends a stamp of confidence to experimental results, as it vastly reduces human error and, consequently, the variability in the data collected,” added Assoc Prof Andreeva.
Thou shall not pass
The data collected from membrane analyses did not sit idle. Instead, it formed the backbone for the ML algorithms implemented. Trained using the data sets churned out by the cobot, the algorithms could be used to explore and predict the properties of GO membranes under various synthesis conditions.
“For example, we could forecast how changes in pH affected the permeability and selectivity of the membranes,” said Assoc Prof Andreeva. “In our study, the recurrent neural network we employed had an accuracy of 88% in predicting whether the membranes would allow ions or water pass through.”
This capability is important as a trained model can tease out the exact parameters required to synthesise a membrane for a specific purpose — offering a smarter route that bypasses the time-consuming and repetitive experiments typically required to achieve a similar effect.
What’s more, there is potential to broaden the automated platform’s capabilities across different research fields. Data amassed from cobot-assisted experiments serves as the building blocks of a universal ML model, which could turn the traditional laboratory practice on its head by enabling more standardised and efficient approaches to material synthesis.
“The concept of ‘self-driving laboratories’ — where cobots plan experiments, execute them and then analyse the results — is exciting,” said Assoc Prof Andreeva. “This could usher in a whole new level of productivity previously unseen in the research laboratory.”