An autonomous laboratory for the accelerated synthesis of novel materials

The study addresses how to accelerate experimental realization of computationally predicted inorganic materials by moving beyond automation to autonomy—enabling an experimental agent to interpret data and make decisions. Prior work has shown autonomy in optimizing carbon nanotube yield, photovoltaic performance and photocatalysis, but effective autonomous synthesis requires fusing encoded domain knowledge, diverse data sources and active learning. The authors present the A-Lab: an autonomous, robotic solid-state synthesis platform that integrates ab initio databases, ML-driven data interpretation, synthesis heuristics learned from text-mined literature, and active learning to plan, execute and optimize syntheses of novel, air-stable inorganic powders at multigram scale.

The paper situates A-Lab within autonomous materials research, citing robotic and Bayesian optimization efforts for carbon nanotubes, photovoltaics and photocatalysis, and autonomous liquid-handling workflows in organic synthesis. It highlights the importance of combining encoded domain knowledge with ML and data resources for autonomy. It leverages large ab initio databases (Materials Project and Google DeepMind) to identify stable or near-stable targets and text-mined literature datasets to learn synthesis analogies (precursor similarity) and heating conditions from prior synthesis reports.

Target selection: Air-stable, previously unreported inorganic compounds were screened from DFT convex hulls using stable phases in the Materials Project and cross-referenced with Google DeepMind; only targets predicted not to react with O2, CO2 and H2O were considered. Each target lies on or near (<10 meV per atom) the convex hull. Recipe generation: Up to five initial recipes per target were proposed by an NLP-based ML model trained on a large, text-mined synthesis database to assess target similarity and select precursors by analogy. A second ML model, trained on literature heating data, proposed synthesis temperatures. Active learning: If initial recipes yielded ≤50% target, the A-Lab used ARROWS, an active-learning algorithm integrating ab initio reaction energies with observed outcomes to predict and prioritize solid-state reaction pathways. It assumes pairwise reaction steps and avoids intermediates that leave small driving forces to reach the target, iterating until majority phase is achieved or search space exhausted. Robotic execution: Three integrated stations—(1) automated powder dosing/mixing into alumina crucibles; (2) robotic loading into one of four box furnaces for heating and cooling; (3) post-heating grinding and XRD characterization—are connected by robotic arms. Operations are controlled via an API enabling autonomous looped operation and human/agent job submission. Characterization and analysis: XRD patterns are analyzed by probabilistic ML models trained on ICSD structures; target patterns are simulated from DFT structures (Materials Project) with corrections to reduce DFT errors. Identified phases and weight fractions are verified by automated Rietveld refinement. Results are reported to a management server to guide subsequent iterations. Knowledge accumulation: The A-Lab builds a database of observed pairwise reactions (88 unique in this campaign). This database allows inference of redundant pathways and pruning of the experimental search space (up to 80% reduction when many precursor sets form identical intermediates). Pathways are prioritized by computed driving forces (Materials Project formation energies), e.g., avoiding low-driving-force intermediates (FePO4, Ca3(PO4)2) to favor higher-driving-force routes (via CaFe3PO8) in synthesizing CaFe2P2O8.

- Through 17 days of continuous, closed-loop operation, A-Lab synthesized 41 of 58 targeted compounds (71% success rate) spanning 33 elements and 41 structural prototypes; 355 total experiments were executed, with 130/355 (≈37%) recipes achieving their targets. - No clear correlation was observed between decomposition energy (relative to the convex hull) and synthesis success across 50 predicted-stable and 8 near-hull metastable targets. - 35/41 successful syntheses were achieved using literature-inspired ML recipes; higher similarity between targets and reference materials increased success likelihood. - Active learning improved yields for nine targets (six with zero initial yield), by avoiding unfavorable pairwise reactions and low-driving-force intermediates; database-driven pruning reduced candidate experiments by up to 80%. - Case study: For CaFe2P2O8, avoiding FePO4 and Ca3(PO4)2 (only 8 meV per atom driving force to target) and routing via CaFe3PO8 increased target yield by ~70%, with a larger driving force (77 meV per atom) at the target-forming step. - Failure analysis identified four primary barriers among 17 unobtained targets: slow reaction kinetics (11 cases, <50 meV per atom driving forces); precursor volatility (e.g., ammonium phosphate evaporation above ~450 °C hindering CaCr2P2O7); product amorphization (e.g., Mo4(PO3)13 forming low-energy amorphous phases ~61 meV per atom above the crystal); and computational inaccuracies (e.g., LaMnO3 mispredicted unstable by ~120 meV per atom; YbMoO4 destabilized by poor pseudopotential; possible lanthanide-related errors for BaGdCrFeO6). - Minor procedural modifications (manual regrinding and higher temperatures) converted two kinetic failures (Y3Ga3In2O12, Mg3NiO4) to successes, projecting an achievable 74% success rate. Excluding targets affected by computational errors projects 78% (43/55) success. - The A-Lab created a reusable database of 88 unique pairwise reactions to inform future syntheses and further reduce search spaces.

The results demonstrate that integrating ab initio screening, literature-derived synthesis heuristics, ML-driven characterization, and active learning in a robotic platform can translate computational predictions into experimental realizations at scale and speed (>2 new materials per day) with minimal human intervention. The high target-level success rate despite modest per-recipe success underscores the importance of intelligent precursor selection and pathway optimization. Failure-mode analysis provides direct guidance for system improvements (e.g., enabling multistep heating, regrinding, broader precursor choices) and reveals how autonomous experiments can feed back to improve computational datasets (identifying DFT and pseudopotential shortcomings, especially for lanthanides and strongly Jahn–Teller-active oxides). The lack of clear correlation between decomposition energy and success highlights the role of kinetics and precursor pathways in determining synthesizability, supporting the use of pathway-aware active learning and pairwise reaction analysis.

The A-Lab showcases an autonomous, modular workflow that accelerates discovery and synthesis of novel inorganic materials by combining DFT databases, text-mined synthesis knowledge, ML-based data interpretation and active learning with robotics. It validated 41 of 58 predicted targets in 17 days and identified actionable strategies to improve outcomes. Future directions include implementing in-line kinetic enhancements (multistep heating, intermittent regrinding), expanding precursor sets beyond air-stable binaries, enriching the pairwise reaction database, improving computational accuracy (e.g., pseudopotentials and treatment of challenging chemistries), and extending the experimental oracle to probe microstructure and device-level performance. Such advancements should further raise success rates and refine our understanding of synthesizability.

- Experimental: Current closed-loop procedures did not include multistep heating, extended dwell times, or intermittent regrinding, limiting ability to overcome slow kinetics; precursor choices were restricted to air-stable binary precursors, sometimes forcing routes through highly stable intermediates; handling in open air may exclude air-/moisture-sensitive targets; melting at high temperatures led to amorphization for some phosphate-rich systems. - Computational: Some targets suffer from DFT inaccuracies (e.g., lanthanides, strong Jahn–Teller systems) and pseudopotential choices, leading to mispredicted stabilities (e.g., LaMnO3, YbMoO4, BaGdCrFeO6). - Generalization: Only powder XRD was used for phase identification; workflow focuses on powder synthesis quantities and may not directly capture microstructure-dependent properties or device performance; the approach was tested on air-stable, near-hull targets and may require adaptation for broader chemistries.