AI-Driven Mobile Robots Revolutionize Chemical Synthesis Research

AI-driven mobile robots team up to tackle chemical synthesis

Researchers at the University of Liverpool have developed AI-driven mobile robots that can carry out chemical synthesis research with axtraordinairy efficiency. In a study publishing in the journal Nature, researchers show how mobile robots that use AI logic to make decisions were able to perform exploratory chemistry research tasks to the same level as humans, but much faster. The 1.75-meter-tall mobile robots were designed by the Liverpool team to tackle three primary problems in exploratory chemistry: performing the reactions, analysing the products, and deciding what to do next based on the data. The two robots performed these tasks in a cooperative manner as they addressed problems in three different areas of chemical synthesis -- structural diversification chemistry (relevant to drug discovery), supramolecular host-guest chemistry, and photochemical synthesis. The results found that with the AI function the mobile robots made the same or similar decisions as a human researcher but these decisions were made on a far quicker timescale than a human, which could take hours. Professor Andrew Cooper from the University of Liverpool's Department of Chemistry and Materials Innovation Factory, who led the project explained: "Chemical synthesis research is time consuming and expensive, both in the physical experiments and the decisions about what experiments to do next so using intelligent robots provides a way to accelerate this process. "When people think about robots and chemistry automation, they tend to think about mixing solutions, heating reactions, and so forth. That's part of it, but the decision making can be at least as time consuming. This is particularly true for exploratory chemistry, where you're not sure of the outcome. It involves subtle, contextual decisions about whether something is interesting or not, based on multiple datasets. It's a time-consuming task for research chemists but a tough problem for AI." Decision-making is a key problem in exploratory chemistry. For example, a researcher might run several trial reactions and then decide to scale up only the ones that give good reaction yields, or interesting products. This is hard for AI to do as the question of whether something is 'interesting' and worth pursuing can have multiple contexts, such as novelty of the reaction product, or the cost and complexity of the synthetic route. Dr Sriram Vijayakrishnan, a former University of Liverpool PhD student and the Postdoctoral Researcher with the Department of Chemistry who led the synthesis work, explained: "When I did my PhD, I did many of the chemical reactions by hand. Often, collecting and figuring out the analytical data took just as long as setting up the experiments. This data analysis problem becomes even more severe when you start to automate the chemistry. You can end up drowning in data." "We tackled this here by building an AI logic for the robots. This processes analytical datasets to make an autonomous decision -- for example, whether to proceed to the next step in the reaction. This decision is basically instantaneous, so if the robot does the analysis at 3:00 am, then it will have decided by 3:01 am which reactions to progress. By contrast, it might take a chemist hours to go through the same datasets." Professor Cooper added: "The robots have less contextual breadth than a trained researcher so in its current form, it won't have a "Eureka!" moment. But for the tasks that we gave it here, the AI logic made more or less the same decisions as a synthetic chemist across these three different chemistry problems, and it makes these decisions in the blink of an eye. There is also huge scope to expand the contextual understanding of the AI, for example by using large language models to link it directly to relevant scientific literature." In the future, the Liverpool team wants to use this technology to discover chemical reactions that are relevant to pharmaceutical drug synthesis, as well as new materials for applications such as carbon dioxide capture. Two mobile robots were used in this study, but there is no limit to the size of the robot teams that could be used. Hence, this approach could scale to the largest industrial laboratories.

AI-driven mobile robots team up to tackle chemical synthesis

Researchers at the University of Liverpool have developed AI-driven mobile robots that can carry out chemical synthesis research with extraordinary efficiency. In a study published in the journal Nature, researchers show how mobile robots that use AI logic to make decisions were able to perform exploratory chemistry research tasks to the same level as humans, but much faster. The paper is titled "Autonomous mobile robots for exploratory synthetic chemistry." The 1.75-meter-tall mobile robots were designed by the Liverpool team to tackle three primary problems in exploratory chemistry: performing the reactions, analyzing the products, and deciding what to do next based on the data. The two robots performed these tasks in a cooperative manner as they addressed problems in three different areas of chemical synthesis -- structural diversification chemistry (relevant to drug discovery), supramolecular host-guest chemistry, and photochemical synthesis. The results found that with the AI function, the mobile robots made the same or similar decisions as a human researcher but these decisions were made on a far quicker timescale than a human, which could take hours. Professor Andrew Cooper from the University of Liverpool's Department of Chemistry and Materials Innovation Factory, who led the project explained, "Chemical synthesis research is time consuming and expensive, both in the physical experiments and the decisions about what experiments to do next, so using intelligent robots provides a way to accelerate this process. "When people think about robots and chemistry automation, they tend to think about mixing solutions, heating reactions, and so forth. That's part of it, but the decision making can be at least as time-consuming. "This is particularly true for exploratory chemistry, where you're not sure of the outcome. It involves subtle, contextual decisions about whether something is interesting or not, based on multiple datasets. It's a time-consuming task for research chemists but a tough problem for AI." Decision-making is a key problem in exploratory chemistry. For example, a researcher might run several trial reactions and then decide to scale up only the ones that give good reaction yields, or interesting products. This is hard for AI to do as the question of whether something is "interesting" and worth pursuing can have multiple contexts, such as novelty of the reaction product, or the cost and complexity of the synthetic route. Dr. Sriram Vijayakrishnan, a former University of Liverpool Ph.D. student and the Postdoctoral Researcher with the Department of Chemistry who led the synthesis work, explained, "When I did my Ph.D., I did many of the chemical reactions by hand. Often, collecting and figuring out the analytical data took just as long as setting up the experiments. This data analysis problem becomes even more severe when you start to automate the chemistry. You can end up drowning in data." "We tackled this here by building an AI logic for the robots. This processes analytical datasets to make an autonomous decision -- for example, whether to proceed to the next step in the reaction. This decision is basically instantaneous, so if the robot does the analysis at 3:00 am, then it will have decided by 3:01 am which reactions to progress. By contrast, it might take a chemist hours to go through the same datasets." Professor Cooper added, "The robots have less contextual breadth than a trained researcher, so in its current form, it won't have a 'Eureka!' moment. But for the tasks that we gave it here, the AI logic made more or less the same decisions as a synthetic chemist across these three different chemistry problems, and it makes these decisions in the blink of an eye. "There is also huge scope to expand the contextual understanding of the AI, for example, by using large language models to link it directly to relevant scientific literature." In the future, the Liverpool team wants to use this technology to discover chemical reactions that are relevant to pharmaceutical drug synthesis, as well as new materials for applications such as carbon dioxide capture. Two mobile robots were used in this study, but there is no limit to the size of the robot teams that could be used. Hence, this approach could scale to the largest industrial laboratories. This new research builds on the world's first "mobile robotic chemist," reported by Professor Cooper's team in 2020 (Nature), which performed almost 700 catalysis experiments over eight days, working 24/7.

Robots with AI brains perform chemical research faster than humans

According to the team at the University of Liverpool, their AI allowed them to make decisions similar to human researchers but much more quickly -- cutting hours of work into a fraction of the time. Chemical synthesis research is time-consuming and costly. It involves physical experiments and decisions about what experiments to conduct next. Intelligent robots offer a solution to accelerate this process. While many associate robots with tasks like mixing solutions or heating reactions, the decision-making aspect can be equally time-consuming. This is especially true in exploratory chemistry, where outcomes are uncertain, and researchers must make nuanced decisions based on multiple datasets. Although challenging for AI, this task is particularly demanding for human chemists. According to researchers, one of the main issues in exploratory chemistry is decision-making. A researcher might, for instance, conduct a number of trial reactions before deciding to scale up only those that produce intriguing compounds or good reaction yields.

Autonomous mobile robots for exploratory synthetic chemistry - Nature

Autonomous robotic laboratories have the potential to change our approach to chemical synthesis, but there are barriers to their widescale adoption. Autonomy implies more than automation; it requires agents, algorithms or artificial intelligence to record and interpret analytical data and to make decisions based on them. This is the key distinction between automated experiments, where the researchers make the decisions, and autonomous experiments, where this is done by machines. The efficacy of autonomous experiments hinges on both the quality and the diversity of the analytical data inputs and their subsequent autonomous interpretation. Automating the decision-making steps in exploratory synthesis is challenging because, unlike some areas of catalysis, it rarely involves the measurement and maximization of a single figure of merit. For example, supramolecular syntheses can produce a wide range of possible self-assembled reaction products, presenting a more open-ended problem from an automation perspective than maximizing the yield of a single, known target. Exploratory synthesis lends itself less well to closed-loop optimization strategies, at least in the absence of a simple quantitative 'novelty' or 'importance' metric. In manual exploratory synthesis, reactions are usually characterized by more than one technique to allow the unambiguous identification of the chemical species. For example, in small-molecule organic syntheses and supramolecular chemistry, mass spectrometry (MS) and nuclear magnetic resonance (NMR) spectroscopy are often combined to probe molecular weight and molecular structure, respectively. Automating the analysis of such multimodal analytical data to guide synthetic discovery processes is not trivial. Artificial-intelligence-based approaches, confined by their training data, might impede genuinely new discoveries by adhering too closely to established prior knowledge. Likewise, rule-based decision methods require careful implementation lest they overlook chemistry that deviates from the rules. More fundamentally, synthetic diversity leads to diverse characterization data. For example, some products in a library might yield highly complex NMR spectra but simple mass spectra, whereas other compounds may show the reverse behaviour, or perhaps give no mass signals at all. As chemists, we make routine, context-based decisions about which data streams to focus on, but this is a major hurdle for autonomous systems. Much progress has been made towards diversifying automated synthesis platforms and increasing their autonomous capabilities. So far, most platforms use bespoke engineering and physically integrated analytical equipment. The associated cost, complexity and proximal monopolization of analytical equipment means that single, fixed characterization techniques are often favoured in automated workflows, rather than drawing on the wider array of analytical techniques available in most synthetic laboratories. This forces any decision-making algorithms to operate with limited analytical information, unlike more multifaceted manual approaches. Hence, closed-loop autonomous chemical synthesis often bears little resemblance to human experimentation, either in the laboratory infrastructure required or in the decision-making steps. We showed previously that free-roaming mobile robots could be integrated into existing laboratories to perform experiments by emulating the physical operations of human scientists. However, that first workflow was limited to one specific type of chemistry -- photochemical hydrogen evolution -- and the only measurement available was gas chromatography, which gives a simple scalar output. Subsequent studies involving mobile robots also focused on the optimization of catalyst performance. These benchtop catalysis workflows cannot carry out more general synthetic chemistry, for example, involving organic solvents, nor can they measure and interpret more complex characterization data, such as NMR spectra. The algorithmic decision-making was limited to maximizing catalyst performance, which is analogous to autonomous synthesis platforms that maximize yield for a reaction using NMR or chromatographic peak areas. Here we present a modular autonomous platform for general exploratory synthetic chemistry. It uses mobile robots to operate a Chemspeed ISynth synthesis platform, an ultrahigh-performance liquid chromatography-mass spectrometer (UPLC-MS) and a benchtop NMR spectrometer. This modular laboratory workflow is inherently expandable to include other equipment, as shown here by the addition of a standard commercial photoreactor. To tackle a broad range of chemistry targets, a heuristic decision-maker was developed to process orthogonal NMR and UPLC-MS data, thus autonomously selecting successful reactions for further study without any human input. This decision-maker also checked the reproducibility of any hits from reaction screens before scale-up. This synthesis-analysis-decision cycle mimics human protocols to make autonomous decisions on the subsequent workflow steps. We exemplify the approach through structural diversification chemistry and the autonomous identification of supramolecular host-guest assemblies. Although the syntheses were autonomous, the choice of chemistry was not: the reactions and building blocks were selected by domain experts before the experiments. This nonetheless gave a large reaction space for the decision-maker to navigate. We also extended this autonomous approach beyond synthesis to assay function by autonomously assessing the host-guest binding properties of successful supramolecular syntheses.