Autonomous mobile robots for exploratory synthetic chemistry

The work addresses how to achieve truly autonomous experimentation in exploratory synthetic chemistry, where outcomes are not easily reduced to a single scalar metric like yield or activity. Existing autonomous laboratories often rely on bespoke, physically integrated equipment and single-mode characterization, limiting decision-making to narrow data streams unlike human practice. Exploratory synthesis, including supramolecular assembly, can yield diverse products and complex mixtures, making closed-loop optimization difficult without simple figures of merit. The study proposes integrating standard laboratory instruments via mobile robots and coupling orthogonal analyses (UPLC–MS and 1H NMR) with heuristic, human-like decision-making to emulate how chemists make context-driven choices across multiple data modalities.

Prior efforts in automated and autonomous synthesis have expanded platform capabilities but frequently center on bespoke hardware and single, integrated analytical methods, constraining decision algorithms to limited data. Earlier mobile-robot studies demonstrated free-roaming robots performing benchtop catalysis with simple scalar outputs (for example, gas chromatography), focusing on catalyst performance optimization. Such systems could not generalize to broader synthetic chemistry involving organic solvents or interpret complex data like NMR. Density/blind optimization approaches, effective when targeting a single response variable, struggle when reactions have multiple possible outcomes or complex mixtures, as in supramolecular chemistry. These limitations motivate more flexible, multimodal, and modular autonomous workflows that mirror human experimentation by combining orthogonal techniques and heuristic decision rules.

The platform is modular, partitioning synthesis and analysis into separate physical modules linked by free-roaming mobile robots for sample transport and handling. Core instruments include a commercial ChemSpeed automated synthesis system, ultrahigh-performance liquid chromatography–mass spectrometry (UPLC–MS), and a benchtop 1H NMR spectrometer; additional modules (for example, a standalone photoreactor) can be appended. Robots operate existing lab equipment without monopolizing it, enabling shared human–robot access. Data orchestration and workflow control are managed via cloud-based software.

Decision-making is heuristic and chemist-designed. After each synthesis–analysis cycle, the decision-maker processes UPLC–MS and 1H NMR data, issuing binary pass/fail outcomes per modality using experiment-specific criteria. These are combined to decide next steps (for example, replicate, scale-up, or diversify reactions). For medicinal chemistry workflows, 1H NMR spectra of reaction mixtures are compared to combined starting-material spectra via dynamic time warping as a distance metric to detect chemical change. UPLC–MS analysis annotates chromatographic peaks, extracts corresponding mass spectra, and matches observed m/z values to pre-generated, plausible combinations derived from known reagents and anticipated products.

Two task-specific modes were explored alongside heuristics to assess scalability. To reduce redundancy, a single robotic module with a multipurpose gripper was shown capable of running two workflows. The overall approach remains expandable to additional instruments and reaction types, including photochemical synthesis, by inserting a remote photoreactor station into the robot-mediated sample flow. For supramolecular discovery, heuristics were deliberately loose to encourage novelty: selections favored samples whose 1H NMR peak counts were comparable to the total for starting materials (consistent with symmetric assemblies), while requiring independent confirmation by plausible m/z in MS. All data, including fails, were archived for reference.

• Demonstrated an end-to-end, autonomous, diverging multi-step medicinal chemistry workflow: initial parallel condensations (acrylates and thioureas) were screened by UPLC–MS and 1H NMR; successful reactions were autonomously scaled up and diversified via two orthogonal strategies (Sonogashira coupling and CuAAC), without intermediate human intervention beyond restocking. Autonomous operation spanned approximately 4 days in this example.
• The decision-maker, using dynamic time warping on 1H NMR and m/z matching for UPLC–MS, correctly identified successful reactions for scale-up and diversification. Manual post hoc inspection confirmed that the software made essentially the same choices a medicinal chemist would.
• An unexpected intramolecular cyclization of one intermediate (same molecular weight as the uncyclized product) was detected by NMR and confirmed by single-crystal X-ray diffraction; it was indistinguishable by UPLC–MS alone, underscoring the need for orthogonal characterization.
• CuAAC reactions produced four of five target molecules (labeled 19–22), validating multi-step autonomous synthesis and heuristic selection.
• Autonomous supramolecular discovery: from 18 combinations of carbonyl-bearing pyridines and amines with metal ions, the decision-maker advanced 2 assemblies (including known cages of the form [Zn2(Ax)2]) based on loose NMR criteria plus MS confirmation. Replicate runs enabled host–guest assays: three guests bound to the cage [Zn2(A2)2] (identified by 1H NMR chemical-shift changes with line broadening), whereas no guest bound to a helicate [Zn2(A9)2], consistent with lack of an internal cavity.
• Modular photochemistry: integrating a remote photoreactor enabled autonomous photocatalyst screening for a decarboxylative conjugate addition. Three photocatalysts yielded the desired product based on UPLC–MS, whereas others (including eosin Y, graphitic carbon nitride, a pyridinium salt) and a blank produced only starting materials. This autonomous photochemical campaign ran over approximately 2 days.
• The platform shared existing lab equipment via mobile robots and handled complex, multimodal data, bringing autonomous decision-making closer to human laboratory practice.

By integrating mobile robots with standard synthesis and analytical instruments and employing heuristic, chemist-informed decision rules over orthogonal data (UPLC–MS and 1H NMR), the platform addresses key barriers to autonomous exploratory synthesis. It moves beyond single-metric optimization to support open-ended discovery, as shown by multi-step medicinal workflows, supramolecular assembly discovery, and function testing (host–guest binding). The detection of an unforeseen cyclization solely via NMR illustrates how multimodal analysis mitigates blind spots inherent to any single technique. The modular, distributed design enables sharing of instruments with human users and straightforward expansion (for example, addition of a photoreactor), making autonomy more accessible and representative of real laboratory practice. Overall, findings indicate that heuristic, human-like decision-making over diverse data streams can effectively guide autonomous exploratory chemistry, even when outcomes are multiple or unanticipated.

The study presents a scalable, modular strategy for autonomous exploratory synthetic chemistry that uses mobile robots to integrate distributed synthesis and analysis platforms. The system autonomously executes screening, scale-up, and diversification; discovers supramolecular assemblies; performs basic function assays; and screens photocatalysts, all using orthogonal UPLC–MS and 1H NMR data with heuristic decision-making. Although not fully closed-loop in the machine-learning sense, the approach significantly narrows the gap between automated and human-like experimentation. Future work could focus on: expanding to additional analytical modalities (for example, higher-field NMR, IR, Raman), improving decision-making via probabilistic or hybrid AI–heuristic methods while preserving openness to novelty, enhancing robustness against edge cases, increasing parallelization and throughput, and further standardizing interfaces to facilitate broader adoption across laboratories.

• Decision-making relied on heuristic rules and pre-generated mass look-up combinations; while enabling openness to novelty, this can yield false positives in single modalities and may miss unanticipated chemistries if not corroborated by orthogonal data.
• Low-field benchtop NMR (60 MHz) produced high dispersion and overlapping signals, potentially inflating apparent peak counts and complicating analyses; criteria were adjusted and cross-validated with MS to compensate.
• Some unexpected edge cases (for example, intramolecular cyclization with indistinguishable m/z and chromatographic behavior) highlight that single-technique assessments can fail; robust decisions required combined modalities.
• Workflows are not fully closed loop in the sense of continuous, model-driven optimization with minimal human-defined criteria; humans still define success heuristics per campaign.
• The approach depends on availability and reliability of multiple instruments and mobile-robot logistics; throughput can be limited by instrument queueing and laboratory space constraints.