Autonomous catalysis research with human–AI–robot collaboration
Negin Orouji, Jeffrey A. Bennett, Richard B. Canty, Long Qi, Shijing Sun, Paulami Majumdar, Chong Liu, Núria López, Neil M. Schweitzer, John R. Kitchin, Hongliang Xin and Milad Abolhasani, “Autonomous catalysis research with human–AI–robot collaboration”, Nature Catalysis 8, 1135–1145 (2025). https://doi.org/10.1038/s41929-025-01430-6
This summary was written by a language model from the paper and has not been checked line by line. Rely on the published paper for details.
This is a Perspective, not a research article: it reports no new experiments or data. The numbers below are the authors’ summaries of results from the studies they review.
Question
Catalyst discovery is still dominated by slow, fragmented, human-centred trial and error, and the available data are fit for narrow purposes but lack negative results, reproducibility and complete synthesis metadata. The authors ask how self-driving laboratories (SDLs), which couple automated synthesis and testing, data infrastructure and AI decision-making in a closed loop, can accelerate catalysis research, which problems are specific to catalysis, and where humans must stay involved (Orouji et al. 2025). Their central position is that SDLs should augment rather than replace researchers, and that real practice spans a spectrum of semi-autonomous workflows rather than a binary of manual versus fully autonomous.
Approach
The Perspective reviews three components of catalysis SDLs: reliable high-throughput automation, data infrastructure with virtual screening, and AI-assisted decision-making. It then argues for an explicit human-in-the-loop role, surveys nine representative studies (summarized in a comparison table of domain, automation type, limitations and links between studies), each chosen to illustrate one of four SDL aspects (hardware, data-driven planning, FAIR data infrastructure, LLM integration), and closes with challenges and opportunities.
Key findings
On methods, the authors contrast predictive models (for example graph neural networks trained on the OC20 dataset of over 1.2 million DFT calculations) that rank existing candidates with generative models that propose candidates to be filtered by physical or data constraints, and suggest coupling them in a loop of generation, predictive triage, automated experiment, data curation and model update. They recommend LLMs as interfaces, planners and retrieval assistants rather than predictors of new chemistry. Bayesian optimization (active learning) suits discrete, costly, independent experiments, while reinforcement learning suits sequential control where the trajectory matters (for example, varying gas feed ratios in CO2 hydrogenation); hybrids and safe variants that enforce pressure or temperature limits are emerging.
For human oversight, they propose three nested validation steps before a discovery claim: in silico cross-checks with physics-informed models, automated replicate runs to measure noise, and a final manual audit. They cite a re-examination of an autonomous synthesis dataset in which apparent hits proved to be artefacts.
The reviewed examples include: BO with parallel fixed-bed reactors that tested 144 catalysts in 42 days for CO2 hydrogenation to methanol, finding a Cu–Zn–Ce/ZrO2 catalyst with 5.7-fold higher CO2 conversion and 12.6-fold higher methanol rate than the best initial catalyst; active-motif ML screening of 465 binary alloys that led to Cu–Pd (32.3% Faradaic efficiency to ethylene) and Cu–Ga (38.4% to formate); DFT plus active learning over about 228,000 sites that led to a de-alloyed Cu–Al catalyst with 80% ethylene Faradaic efficiency at 400 mA cm-2 for over 100 h; BO with manual experiments for higher alcohol synthesis that found Fe65Co19Cu5Zr11 (1.1 g h-1 gcat-1, 150 h stable) in 86 to 104 experiments from about 5 billion possibilities; the Swiss CAT+ FAIR data platform (up to 96 batch or 16 flow reactors, 100 bar, 925 °C); a mobile robot chemist for photocatalysis; Fast-Cat, which mapped the yield–regioselectivity Pareto front of rhodium-catalysed 1-octene hydroformylation in 45 experiments over five days; Coscientist and the LLM-RDF framework (94.5% aldehyde yield in aerobic alcohol oxidation); and SAMPLE, whose four agents found enzyme variants at least 12 °C more thermostable while sampling under 2% of a 1,352-member landscape.
The challenges they list are delayed experimental validation, scarce structured data and metadata, weak integration of synthesis and testing, and few high-throughput in situ measurements. Heterogeneous catalysis is singled out as harder to automate because of powder handling, high pressures and temperatures, and operando probes that do not fit robotic workflows.
Significance
The paper gives a catalysis-specific roadmap for autonomous experimentation, with four opportunities: hybrid computational–experimental workflows (models pretrained on simulations and refined with experiments), ML-identified proxy measurements, AI-guided co-design of synthesis and testing, and open science practices for sharing datasets, robotic protocols and models (data sharing, reproducibility). Its emphasis on validation and human judgement complements broader SDL perspectives from overlapping authors.