Roadmap for transforming heterogeneous catalysis with artificial intelligence
Hongliang Xin, John R. Kitchin, Núria López, Neil M. Schweitzer, Nongnuch Artrith, Fanglin Che, Lars C. Grabow, G. T. Kasun Kalhara Gunasooriya, Heather J. Kulik, Teodoro Laino, Hao Li, Suljo Linic, Andrew J. Medford, Randall J. Meyer, Jiayu Peng, Cory Phillips, Jin Qian, Long Qi, Wendy J. Shaw, Zachary W. Ulissi, Siwen Wang and Xiaonan Wang, “Roadmap for transforming heterogeneous catalysis with artificial intelligence”, Nature Catalysis 9(2), 102–111 (2026). https://doi.org/10.1038/s41929-026-01479-x
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 written by a large community of authors (its acknowledgements credit an NSF-funded workshop), so it reports no new data; the summary describes its arguments and the examples it cites.
Question
The authors ask what it would take for artificial intelligence to move heterogeneous catalysis beyond trial and error, given the obstacles they see as most serious: scarce and poorly reproducible data, deep learning models that fail to generalize outside their training domain, a lack of mechanistically grounded interpretation of dynamically evolving active sites, and the missing closed loop between computation and experiment (Xin et al. 2026).
Approach
The Perspective lays out a roadmap with three linked pillars that build on one another: an AI-ready data ecosystem, multimodal foundation models, and autonomous laboratories, culminating in what the authors call agentic catalysis through AI-empowered human-machine collaboration. For each pillar it surveys current tools and examples, names failure modes, and proposes guardrails and progress metrics. It closes with a picture of networked cloud laboratories and a call to funders, industry and policymakers.
Key findings
Data ecosystem. Catalysis data are scattered across papers, notebooks and disconnected repositories, often without ontologies or metadata, and negative or null results are rarely reported, which biases models. The authors argue for FAIR (findable, accessible, interoperable, reusable) data underpinned by domain ontologies, electronic lab notebooks and laboratory information management systems (Chemotion with its LabIMotion extension is given as an example), and computational tools with provenance such as ASE, pymatgen, FireWorks and AiiDA. They point to Catalysis-Hub, the Open Catalyst datasets (OC20, OC22, OMat24) that led toward universal models for atoms, ioChem-BD, NOMAD, DigCat and the CatTestHub benchmark as building blocks. Rather than choosing between structured databases and large language models, they suggest investing in structured, reproducible data for benchmarking while using LLMs for mining and curation. Proposed progress metrics include inter-laboratory reproducibility thresholds, metadata completeness, time from measurement to FAIR release, and the reduction in model error when training on standardized rather than siloed data. Industry participation is singled out as a major hurdle.
Multimodal foundation models. The authors describe five learning strategies such models would combine. Representation learning (graph neural networks, physics-informed networks such as TinNet, attention-based graph Transformers) learns features from raw data. Transfer learning addresses data scarcity; one cited study transferred reactivity knowledge between photocatalytic reactions using as few as ten new experimental points. Generative models (VAEs, GANs, diffusion and flow models) have proposed Rh-Ru alloy motifs for ammonia synthesis and Cu-Al and Sn-Pd alloys for CO2 reduction that were later validated. Contrastive learning aligns modalities; a graph-text model of adsorbate-surface configurations cut adsorption energy error by about 10% relative to a text-only model. Reinforcement learning handles sequential decisions, for example discovering a lower-barrier NH3 pathway on Fe(111) and the CatGym framework for surface segregation. For each strategy the paper lists risks (proxy correlations, negative transfer and catastrophic forgetting, unphysical or unsynthesizable proposals, batch effects, unsafe exploration) and guardrails such as uncertainty quantification, out-of-distribution testing, stability filters, model and data cards, multimodal benchmarks, and time-split validation.
Autonomous laboratories. The authors describe five stages: task automation, workflow automation, intelligent planning, closed-loop automation, and AI-empowered human-machine collaboration. As an example of current capability they cite a Bayesian optimization platform that made and tested 144 CO2-to-methanol catalyst formulations in six weeks, while stressing that such systems are still rule-based and tool-centric rather than goal-setting. Hardware is not yet versatile enough and the orchestration software is fragmented, so they expect a continuum of increasingly agentic laboratories rather than a sudden leap.
Networked agentic catalysis. The longer-term picture is a shared cloud network where each laboratory runs programmable workflows linking catalyst identity, operando conditions, performance and characterization, with ontology-tagged, versioned records that include uncertainties and negative results. A planner-executor scheme would check actions against a digital twin that includes deactivation and regeneration. It would use Bayesian optimization for single-step choices and model-based or offline reinforcement learning for multi-step sequences, with a guardrail stack that blocks high-uncertainty proposals for human review and requires replication and shadow-mode testing before anything is adopted.
Significance
The paper offers a community consensus agenda for AI in catalysis that treats data infrastructure, models and autonomy as one connected program, and it gives concrete safeguards and metrics rather than only aspirations. It also argues that education, incentives for sharing data (including negative results), ethical governance, federated learning and sustained funding are needed to keep progress from fragmenting into proprietary silos (Xin et al. 2026).