pith. sign in

arxiv: 2606.05050 · v1 · pith:JIDVP7F6new · submitted 2026-06-03 · ❄️ cond-mat.mtrl-sci · cs.AI· physics.chem-ph

Autonomous heterogeneous catalyst discovery with a self-evolving multi-agent digital twin

Zhilong Song , Zongmin Zhang , Lixue Cheng This is my paper

Pith reviewed 2026-06-28 05:08 UTC · model grok-4.3

classification ❄️ cond-mat.mtrl-sci cs.AIphysics.chem-ph
keywords heterogeneous catalysisdigital twinmulti-agent systemcatalyst discoverypropane dehydrogenationreaction pathway searchsurface reconstructionturnover frequency prediction
0
0 comments X

The pith

CatDT builds a self-evolving multi-agent digital twin that predicts heterogeneous catalyst kinetics within 0.5-2 times experimental values across seven benchmarks spanning four orders of magnitude.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces CatDT as an autonomous system that starts from a bulk crystal structure and a natural-language reaction description to model working catalyst surfaces and compute reaction rates. It establishes that eight specialized agents combined with 27 deterministic tools can handle facet prediction, surface reconstruction, pathway enumeration, and kinetics calculation in minutes on a single GPU. The central demonstration is that two engineered components, UniMech for low-cost dominant-pathway identification and a memory-augmented reinforcement loop for transition-state calculations, produce results that stay inside experimental scatter for stepped metals, single-atom catalysts, intermetallics, 2D materials, and SMSI interfaces. The work further shows the system can propose non-precious candidates for propane dehydrogenation whose simulated turnover frequencies and selectivities approach those of the industrial platinum benchmark.

Core claim

From only a bulk crystal and reaction description, the CatDT multi-agent system reconstructs active surfaces, identifies dominant reaction pathways at over 1000 times lower cost than exhaustive search via UniMech, raises barrier-calculation success from 41 percent to 84 percent through memory-augmented reinforcement, and delivers turnover-frequency predictions lying between 0.5 and 2 times experiment on seven gas-solid test cases; on propane dehydrogenation it nominates a Ni@ZrO2 SMSI overlayer with a simulated TOF of 1.63 s^{-1} at near-100 percent selectivity, matching or exceeding the Pt benchmark.

What carries the argument

CatDT, the self-evolving multi-agent digital twin whose UniMech module fuses agent-guided proposals with energy-cached graph search to locate dominant pathways and whose memory-augmented reinforcement loop improves transition-state success rates.

If this is right

  • Catalyst screening for arbitrary gas-solid and liquid-solid reactions can proceed from crystal structure alone without manual pathway enumeration.
  • Non-precious metal and oxide combinations can be ranked for activity and selectivity before synthesis.
  • The same agent-and-tool harness can be applied to other multi-stage scientific simulators that require repeated energy evaluations.
  • Persistent memory across runs allows the system to improve on repeated tasks without retraining the underlying language model.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The reported success on SMSI interfaces suggests the framework could be extended to other buried or reconstructed interfaces once the surface-reconstruction agent is generalized.
  • If the 5-30 minute runtime holds for larger unit cells, high-throughput screening of ternary and quaternary compositions becomes feasible on modest hardware.
  • The emphasis on deterministic tools over raw model scale implies similar harnesses could improve reliability in adjacent domains such as battery electrolyte design or enzyme mechanism prediction.

Load-bearing premise

The combination of LLM-guided agent proposals, deterministic tools, and memory-augmented loops will locate the true dominant pathways and deliver accurate kinetics without systematic omissions or biases for previously unseen materials.

What would settle it

Laboratory measurement of turnover frequency and product selectivity for the proposed Ni@ZrO2 SMSI overlayer under propane dehydrogenation conditions at the temperature and pressure used in the simulation.

read the original abstract

Theoretical heterogeneous catalysis promises rapid catalyst discovery, yet computational and machine-learning predictions often deviate from experiment and stay confined to narrow material families, for want of a faithful, condition-aware catalytic simulator. We present CatDT (Catalysis Digital Twin), a self-evolving multi-agent system that builds an autonomous digital twin of a working catalyst, unifying gas-solid and liquid-solid modeling. From only a bulk crystal and a natural-language reaction description, eight specialized agents and 27 scientific tools predict stable facets, reconstruct working surfaces, enumerate and rank reaction pathways, locate transition states, and compute kinetics in 5-30 min on a single GPU. Two innovations address the hardest steps: UniMech finds dominant pathways for novel materials at over $10^3\times$ lower cost than exhaustive enumeration by fusing agent-guided proposals with energy-cached graph search, and a memory-augmented reinforcement loop raises barrier-calculation success from 41\% to 84\% across 600 catalytic surfaces. Across seven gas-solid benchmarks -- stepped metals, single-atom catalysts, ordered intermetallics, vacancy-rich 2D sulfides and carbides, and a strong-metal--support-interaction (SMSI) interface -- every CatDT prediction lies within 0.5-2 times experiment over four orders of magnitude. For propane dehydrogenation, CatDT independently discovers non-precious candidates rivaling the Pt-based industrial benchmark, with a proposed Ni@ZrO$_2$ SMSI overlayer reaching a simulated TOF of $1.63~\text{s}^{-1}$ at $\sim$100\% selectivity. More broadly, the decisive factor for a faithful catalyst digital twin -- or any multi-stage scientific simulator -- is not raw LLM capability but the engineered harness around it: deterministic tools, persistent memory, and verified self-improvement that compound across models, tools, and runs.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The paper presents CatDT, a self-evolving multi-agent digital twin for autonomous heterogeneous catalyst discovery. From only a bulk crystal structure and natural-language reaction description, eight specialized agents and 27 tools predict stable facets, reconstruct working surfaces, enumerate and rank pathways via the UniMech module (agent-guided proposals fused with energy-cached graph search), locate transition states with a memory-augmented reinforcement loop, and compute condition-aware kinetics in 5-30 min on one GPU. The central empirical claim is that across seven gas-solid benchmarks (stepped metals, SACs, intermetallics, 2D sulfides/carbides, SMSI interfaces) every predicted TOF lies within 0.5-2 imes of experiment over four orders of magnitude; the system is further applied to propane dehydrogenation, where it proposes a non-precious Ni@ZrO2 SMSI overlayer with simulated TOF 1.63 s^{-1} at ~100% selectivity.

Significance. If the reported experimental agreement and pathway completeness hold, the work would constitute a notable advance toward general, condition-aware simulators in computational catalysis, moving beyond narrow material families and manual pathway enumeration. The engineered combination of deterministic tools, persistent memory, and self-improving reinforcement (raising TS success from 41% to 84%) directly addresses a recognized bottleneck; the >10^3 cost reduction claimed for UniMech and the discovery of a competitive non-precious candidate are concrete strengths that, if independently reproducible, could influence future multi-agent scientific simulators.

major comments (3)
  1. [Abstract / Methods (UniMech)] Abstract and Methods (UniMech description): the headline claim that every CatDT prediction lies within 0.5-2 imes experiment across the seven benchmarks is load-bearing for the central thesis, yet the manuscript provides no explicit audit or completeness test showing that the eight agents did not omit low-energy intermediates or elementary steps for the novel or interface systems (e.g., the Ni@ZrO2 SMSI case); because UniMech cannot recover routes absent from the agent proposals, the reported agreement may be conditional on correct a-priori proposal rather than general pathway discovery.
  2. [Results (benchmarks)] Results (seven gas-solid benchmarks): the selection criteria, exclusion rules, and prior knowledge of dominant pathways for the seven benchmarks are not stated; without this information it is impossible to assess whether the 0.5-2 imes agreement generalizes or reflects post-hoc filtering of cases where the agent-guided search succeeded.
  3. [Methods (reinforcement loop)] Methods (memory-augmented reinforcement loop): while the loop improves TS-finding success from 41% to 84% across 600 surfaces, the manuscript does not clarify whether this improvement propagates to the final pathway ranking or TOF values, or whether failed TS searches are handled by fallback assumptions that could bias the reported kinetics.
minor comments (2)
  1. [Abstract] Abstract: the acronym 'SMSI' is used before its expansion; a parenthetical on first use would improve readability for a broad audience.
  2. [Abstract] Abstract: the phrase 'over four orders of magnitude' is stated without an accompanying table or figure reference listing the experimental TOF range; adding such a pointer would strengthen the quantitative claim.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed comments, which help strengthen the presentation of our work. We respond to each major comment below and indicate where revisions to the manuscript are warranted.

read point-by-point responses

  1. Referee: [Abstract / Methods (UniMech)] Abstract and Methods (UniMech description): the headline claim that every CatDT prediction lies within 0.5-2 imes experiment across the seven benchmarks is load-bearing for the central thesis, yet the manuscript provides no explicit audit or completeness test showing that the eight agents did not omit low-energy intermediates or elementary steps for the novel or interface systems (e.g., the Ni@ZrO2 SMSI case); because UniMech cannot recover routes absent from the agent proposals, the reported agreement may be conditional on correct a-priori proposal rather than general pathway discovery.

    Authors: We agree that an explicit completeness audit is not provided in the current manuscript and that this is a substantive point given UniMech's dependence on agent proposals. The eight agents encode chemical heuristics via the 27 tools to generate proposals, after which the energy-cached graph search ranks and expands within the proposed space; however, routes outside the initial proposals cannot be recovered. For the seven benchmarks we relied on literature-established dominant pathways, but for the Ni@ZrO2 SMSI case no such external validation of completeness exists. We will add a dedicated subsection in Methods (and supplementary material) that (i) describes the proposal-generation rules, (ii) reports an internal audit comparing CatDT pathways against all literature-reported steps for the benchmark systems, and (iii) explicitly states the limitation for truly novel interfaces. This constitutes a partial revision because exhaustive enumeration remains computationally prohibitive. revision: partial

  2. Referee: [Results (benchmarks)] Results (seven gas-solid benchmarks): the selection criteria, exclusion rules, and prior knowledge of dominant pathways for the seven benchmarks are not stated; without this information it is impossible to assess whether the 0.5-2 imes agreement generalizes or reflects post-hoc filtering of cases where the agent-guided search succeeded.

    Authors: The referee is correct that the manuscript does not explicitly document the benchmark selection process. The seven systems were chosen to span distinct catalyst classes (stepped metals, SACs, intermetallics, 2D materials, SMSI interfaces) for which experimental TOF values spanning four orders of magnitude are available in the literature; systems with known multi-site or coverage-dependent mechanisms outside the current tool set were excluded. We will insert a new subsection in Results that lists the selection criteria, the exclusion rules applied, and the literature sources used to identify the dominant pathways against which CatDT results were compared. This addition will allow readers to evaluate the scope of the reported agreement. revision: yes

  3. Referee: [Methods (reinforcement loop)] Methods (memory-augmented reinforcement loop): while the loop improves TS-finding success from 41% to 84% across 600 surfaces, the manuscript does not clarify whether this improvement propagates to the final pathway ranking or TOF values, or whether failed TS searches are handled by fallback assumptions that could bias the reported kinetics.

    Authors: We acknowledge that the manuscript reports the success-rate improvement but does not trace its effect on the final TOF values or describe fallback procedures for the remaining 16% of cases. In the implementation, failed TS searches trigger additional reinforcement episodes with varied initial conditions; if still unsuccessful, the pathway is either discarded (if it is not rate-limiting) or assigned an upper-bound barrier from a related surface. We will revise the Methods section to (i) quantify how many of the final benchmark pathways benefited from the improved loop, (ii) state the fallback rules explicitly, and (iii) report the sensitivity of the reported TOFs to those fallbacks. This will be a full revision of the relevant subsection. revision: yes

Circularity Check

0 steps flagged

No significant circularity; central claims rest on external experimental benchmarks

full rationale

The paper's load-bearing result is the reported 0.5-2× agreement between CatDT predictions and independent experimental TOFs across seven gas-solid benchmarks spanning multiple material classes. This comparison uses quantities measured outside the model (experimental rates) rather than quantities defined or fitted inside the multi-agent loop. No equations, pathway enumerations, or reinforcement updates are shown to reduce by construction to the same inputs; the UniMech graph search and memory-augmented loop are algorithmic components whose outputs are then tested against external data. No self-citations, ansatzes smuggled via prior work, or renamings of known results appear as load-bearing steps in the derivation chain. The method is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 2 invented entities

The approach rests on the assumption that LLM agents can be reliably harnessed by deterministic tools for catalysis modeling; without the full text, specific free parameters in underlying models or energy calculations cannot be enumerated.

free parameters (1)
  • Agent and tool hyperparameters
    The multi-agent system and reinforcement loop likely contain multiple tunable parameters whose values are not specified in the abstract.
axioms (1)
  • domain assumption LLM-based agents combined with scientific tools can faithfully reconstruct working catalyst surfaces and locate transition states from bulk input alone
    This premise underpins the entire autonomous pipeline described in the abstract.
invented entities (2)
  • UniMech no independent evidence
    purpose: Agent-guided proposal fused with energy-cached graph search for dominant pathway identification at low cost
    New component introduced to address exhaustive enumeration cost.
  • CatDT multi-agent digital twin no independent evidence
    purpose: Unified gas-solid and liquid-solid catalytic simulator
    The overall self-evolving system is the central new construct.

pith-pipeline@v0.9.1-grok · 5880 in / 1461 out tokens · 33268 ms · 2026-06-28T05:08:15.760895+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Socratic agents for autonomous scientific discovery in high-dimensional physical systems

    cs.AI 2026-06 unverdicted novelty 6.0

    AHOIS is a Socratic multi-agent AI that autonomously discovers and validates a random-interference encoding strategy for multimode fiber optics, achieving 76.97% MNIST and 83.17% Fashion-MNIST accuracy with 16x16 meas...

Reference graph

Works this paper leans on

77 extracted references · 4 linked inside Pith · cited by 1 Pith paper

  1. [1]

    K., Bligaard, T., Rossmeisl, J

    Nørskov, J. K., Bligaard, T., Rossmeisl, J. & Christensen, C. H. Towards the computational design of solid catalysts. Nat. Chem. 1, 37–46 (2009)

    2009

  2. [2]

    K., Abild-Pedersen, F., Studt, F

    Nørskov, J. K., Abild-Pedersen, F., Studt, F. & Bligaard, T. Density functional theory in surface chemistry and catalysis. Proc. Natl. Acad. Sci. USA 108, 937–943 (2011)

    2011

  3. [3]

    Zhong, M. et al. Accelerated discovery of CO 2 electrocatalysts using active machine learning. Nature 581, 178–183 (2020)

    2020

  4. [4]

    W., Medford, A

    Ulissi, Z. W., Medford, A. J., Bligaard, T. & Nørskov, J. K. To address surface reaction network complexity using scaling relations machine learning and DFT calculations. Nat. Commun. 8, 14621 (2017)

    2017

  5. [5]

    & J´ onsson, H

    Henkelman, G. & J´ onsson, H. Improved tangent estimate in the nudged elastic band method for finding minimum energy paths and saddle points. J. Chem. Phys. 113, 9978–9985 (2000)

    2000

  6. [6]

    Henkelman, G., Uberuaga, B. P. & J´ onsson, H. A climbing image nudged elastic band method for finding saddle points and minimum energy paths. J. Chem. Phys. 113, 9901–9904 (2000)

    2000

  7. [7]

    & Sprik, M

    Cheng, J., Liu, X., VandeVondele, J., Sulpizi, M. & Sprik, M. Redox potentials and acidity constants from density functional theory based molecular dynamics. Acc. Chem. Res. 47, 3522–3529 (2014)

    2014

  8. [8]

    Logad´ ottir,´A. et al. The Brønsted–Evans–Polanyi relation and the volcano plot for ammonia synthesis over transition metal catalysts. J. Catal. 197, 229–231 (2001)

    2001

  9. [9]

    Bligaard, T. et al. The Brønsted–Evans–Polanyi relation and the volcano curve in heterogeneous catalysis. J. Catal. 224, 206–217 (2004)

    2004

  10. [10]

    Abild-Pedersen, F. et al. Scaling properties of adsorption energies for hydrogen-containing molecules on transition-metal surfaces. Phys. Rev. Lett. 99, 016105 (2007)

    2007

  11. [11]

    Calle-Vallejo, F., Loffreda, D., Koper, M. T. M. & Sautet, P. Introducing structural sensitivity into adsorption-energy scaling relations by means of coordination numbers. Nat. Chem. 7, 403–410 (2015)

    2015

  12. [12]

    T., Sauceda, H

    Sch¨ utt, K. T., Sauceda, H. E., Kindermans, P.-J., Tkatchenko, A. & M¨ uller, K.-R. SchNet – a deep learning architecture for molecules and materials. J. Chem. Phys. 148, 241722 (2018)

    2018

  13. [13]

    Gasteiger, J., Giri, S., Margraf, J. T. & G¨ unnemann, S. Fast and uncertainty-aware directional message passing for non-equilibrium molecules. In NeurIPS 2020 Workshop on Machine Learning for Molecules (2020)

    2020

  14. [14]

    & G¨ unnemann, S

    Gasteiger, J., Becker, F. & G¨ unnemann, S. GemNet: Universal directional graph neural 34 networks for molecules. In Adv. Neural Inf. Process. Syst. , vol. 34, 6790–6802 (2021)

    2021

  15. [15]

    P., Simm, G

    Batatia, I., Kov´ acs, D. P., Simm, G. N. C., Ortner, C. & Cs´ anyi, G. MACE: Higher order equivariant message passing neural networks for fast and accurate force fields. In Adv. Neural Inf. Process. Syst. , vol. 35, 11423–11436 (2022)

    2022

  16. [16]

    & Smidt, T

    Liao, Y.-L., Wood, B., Das, A. & Smidt, T. EquiformerV2: Improved equivariant transformer for scaling to higher-degree representations. In Int. Conf. Learn. Represent. (2024)

    2024

  17. [17]

    Chanussot, L. et al. Open Catalyst 2020 (OC20) dataset and community challenges. ACS Catal. 11, 6059–6072 (2021)

    2020

  18. [18]

    Tran, R. et al. The Open Catalyst 2022 (OC22) dataset and challenges for oxide electrocatalysts. ACS Catal. 13, 3066–3084 (2023)

    2022

  19. [19]

    Sahoo, S. J. et al. The Open Catalyst 2025 (OC25) dataset and models for solid–liquid interfaces. arXiv preprint arXiv:2509.17862 (2025)

    Pith/arXiv arXiv 2025

  20. [20]

    Barroso-Luque, L. et al. UMA: A family of universal models for atoms. arXiv preprint arXiv:2506.23971 (2025)

    arXiv 2025

  21. [21]

    Yin, J., Chen, H., Qiu, J. et al. SurFF: a foundation model for surface exposure and morphology across intermetallic crystals. Nat. Comput. Sci. 5, 782–792 (2025)

    2025

  22. [22]

    Du, X. et al. Machine-learning-accelerated simulations to enable automatic surface reconstruc- tion. Nat. Comput. Sci. 3, 1034–1044 (2023)

    2023

  23. [23]

    Du, X., Damewood, J. K. & G´ omez-Bombarelli, R. Accelerating and enhancing thermodynamic simulations of electrochemical interfaces. ACS Cent. Sci. 11, 1558–1572 (2025)

    2025

  24. [24]

    Lan, J. et al. AdsorbML: a leap in efficiency for adsorption energy calculations using generalizable machine learning potentials. npj Comput. Mater. 9, 172 (2023)

    2023

  25. [25]

    & Kitchin, J

    Kolluru, A. & Kitchin, J. R. AdsorbDiff: Adsorbate placement via conditional denoising diffusion. In Proc. 41st Int. Conf. Mach. Learn. , vol. 235, 25042–25057 (2024)

    2024

  26. [26]

    Morandi, S. et al. An end-to-end framework for reactivity in heterogeneous catalysis. Nat. Chem. Eng. 3, 169–180 (2026)

    2026

  27. [27]

    P., Grimmel, S

    Unsleber, J. P., Grimmel, S. A. & Reiher, M. Chemoton 2.0: Autonomous exploration of chemical reaction networks. J. Chem. Theory Comput. 18, 5393–5409 (2022)

    2022

  28. [28]

    & Savoie, B

    Zhao, Q., Xu, Y., Greeley, J. & Savoie, B. M. Deep reaction network exploration at a heterogeneous catalytic interface. Nat. Commun. 13, 4860 (2022)

    2022

  29. [29]

    R., Zitnick, C

    Wander, B., Shuaibi, M., Kitchin, J. R., Zitnick, C. L. & Ulissi, Z. W. CatTSunami: Accelerating 35 transition state energy calculations with pre-trained graph neural networks. arXiv preprint arXiv:2405.02078 (2024)

    arXiv 2024

  30. [30]

    & Liu, Y

    Wang, R., Fang, S., Huang, Q. & Liu, Y. Constant-potential machine learning force field for the electrochemical interface. J. Chem. Theory Comput. 21, 7628–7635 (2025)

    2025

  31. [31]

    Medford, A. J. et al. CatMAP: A software package for descriptor-based microkinetic mapping of catalytic trends. Catal. Lett. 145, 794–807 (2015)

    2015

  32. [32]

    & Stamatakis, M

    Pineda, M. & Stamatakis, M. Kinetic Monte Carlo simulations for heterogeneous catalysis: A tutorial review of recent developments. J. Chem. Phys. 156, 120902 (2022)

    2022

  33. [33]

    Wellendorff, J. et al. Density functionals for surface science: Exchange-correlation model development with Bayesian error estimation. Phys. Rev. B 85, 235149 (2012)

    2012

  34. [34]

    Wen, M. et al. Chemical reaction networks and opportunities for machine learning. Nat. Comput. Sci. 3, 12–24 (2023)

    2023

  35. [35]

    & Savoie, B

    Zhao, Q. & Savoie, B. M. Simultaneously improving reaction coverage and computational cost in automated reaction prediction tasks. Nat. Comput. Sci. 1, 479–490 (2021)

    2021

  36. [36]

    Mou, T. et al. Bridging the complexity gap in computational heterogeneous catalysis with machine learning. Nat. Catal. 6, 122–136 (2023)

    2023

  37. [37]

    Qian, J. et al. Digital twin for chemical science: a case study on water interactions on the Ag(111) surface. Nat. Comput. Sci. 5, 793–800 (2025)

    2025

  38. [38]

    A., MacKnight, R., Kline, B

    Boiko, D. A., MacKnight, R., Kline, B. & Gomes, G. Autonomous chemical research with large language models. Nature 624, 570–578 (2023)

    2023

  39. [39]

    Bran, A. M. et al. ChemCrow: Augmenting large-language models with chemistry tools. Nat. Mach. Intell. 6, 525–535 (2024)

    2024

  40. [40]

    & Chen, Z

    Li, Y., Zhang, H., Sun, Y., Wang, X. & Chen, Z. Catalyst-Agent: Autonomous heterogeneous catalyst screening and optimization with an LLM agent. arXiv preprint arXiv:2603.01311 (2026)

    Pith/arXiv arXiv 2026

  41. [41]

    Li, G., Hammoud, H. A. A. K., Itani, H., Khizbullin, D. & Ghanem, B. CAMEL: Communicative agents for “mind” exploration of large language model society. In Adv. Neural Inf. Process. Syst., vol. 36, 51991–52008 (2023)

    2023

  42. [42]

    C., Collison, C

    Ramos, M. C., Collison, C. J. & White, A. D. A review of large language models and autonomous agents in chemistry. Chem. Sci. 16, 2514–2572 (2025)

    2025

  43. [43]

    Zhou, C. et al. Externalization in LLM agents: a unified review of memory, skills, protocols 36 and harness engineering. arXiv preprint arXiv:2604.08224 (2026)

    Pith/arXiv arXiv 2026

  44. [44]

    Bui, N. D. Q. Building effective AI coding agents for the terminal: scaffolding, harness, context engineering, and lessons learned. arXiv preprint arXiv:2603.05344 (2026)

    arXiv 2026

  45. [45]

    & Zheng, H.-T

    Pan, L., Zou, L., Guo, S., Ni, J. & Zheng, H.-T. Natural-language agent harnesses. arXiv preprint arXiv:2603.25723 (2026)

    Pith/arXiv arXiv 2026

  46. [46]

    Lou, X. et al. AutoHarness: improving LLM agents by automatically synthesizing a code harness. arXiv preprint arXiv:2603.03329 (2026)

    arXiv 2026

  47. [47]

    Zhou, H. et al. Memento: Fine-tuning LLM agents without fine-tuning LLMs. arXiv preprint arXiv:2508.16153 (2025)

    arXiv 2025

  48. [48]

    & Pidko, E

    Hashemi, A., Bougueroua, S., Gaigeot, M.-P. & Pidko, E. A. ReNeGate: A reaction net- work graph-theoretical tool for automated mechanistic studies in computational homogeneous catalysis. J. Chem. Theory Comput. 18, 7470–7482 (2022)

    2022

  49. [49]

    & Reiher, M

    Steiner, M. & Reiher, M. Autonomous reaction network exploration in homogeneous and heterogeneous catalysis. Top. Catal. 65, 6–39 (2022)

    2022

  50. [50]

    Kreitz, B. et al. Detailed microkinetics for the oxidation of exhaust gas emissions through automated mechanism generation. ACS Catal. 12, 11137–11151 (2022)

    2022

  51. [51]

    A., Markvoort, A

    van Santen, R. A., Markvoort, A. J., Filot, I. A. W., Ghouri, M. M. & Hensen, E. J. M. Mechanism and microkinetics of the Fischer–Tropsch reaction. Phys. Chem. Chem. Phys. 15, 17038–17063 (2013)

    2013

  52. [52]

    & Comas-Vives, A

    Foppa, L., Iannuzzi, M., Cop´ eret, C. & Comas-Vives, A. Facile Fischer–Tropsch chain growth from CH2 monomers enabled by the dynamic CO adlayer. ACS Catal. 9, 6571–6582 (2019)

    2019

  53. [53]

    Honkala, K. et al. Ammonia synthesis from first-principles calculations. Science 307, 555–558 (2005)

    2005

  54. [54]

    Zhao, H. et al. The role of Cu 1–O3 species in single-atom Cu/ZrO 2 catalyst for CO 2 hydro- genation. Nat. Catal. 5, 818–831 (2022)

    2022

  55. [55]

    Studt, F. et al. Discovery of a ni-ga catalyst for carbon dioxide reduction to methanol. Nat. Chem. 6, 320–324 (2014)

    2014

  56. [56]

    & Hosono, H

    Sugiyama, H., Miyazaki, M., Sasase, M., Kitano, M. & Hosono, H. Room-temperature co 2 hydrogenation to methanol over air-stable hcp-pdmo intermetallic catalyst. J. Am. Chem. Soc. 145, 9410–9416 (2023)

    2023

  57. [57]

    Hu, J. et al. Sulfur vacancy-rich MoS2 as a catalyst for the hydrogenation of CO 2 to methanol. 37 Nat. Catal. 4, 242–250 (2021)

    2021

  58. [58]

    Zhou, H. et al. Two-dimensional molybdenum carbide 2D-Mo 2C as a superior catalyst for CO 2 hydrogenation. Nat. Commun. 12, 5510 (2021)

    2021

  59. [59]

    Chen, S. et al. Defective TiOx overlayers catalyze propane dehydrogenation promoted by base metals. Science 385, 295–300 (2024)

    2024

  60. [60]

    H., Almallahi, R., Wortman, J., Igenegbai, V

    Motagamwala, A. H., Almallahi, R., Wortman, J., Igenegbai, V. O. & Linic, S. Stable and selective catalysts for propane dehydrogenation operating at thermodynamic limit. Science 373, 217–222 (2021)

    2021

  61. [61]

    Sun, G. et al. Breaking the scaling relationship via thermally stable Pt/Cu single atom alloys for catalytic dehydrogenation. Nat. Commun. 9, 4454 (2018)

    2018

  62. [62]

    Rochlitz, L. et al. A robust and efficient propane dehydrogenation catalyst from unexpectedly segregated Pt2Mn nanoparticles. J. Am. Chem. Soc. 144, 13384–13393 (2022)

    2022

  63. [63]

    He, Y., Song, Y., Cullen, D. A. & Laursen, S. Selective and stable non-noble-metal intermetallic compound catalyst for the direct dehydrogenation of propane to propylene. J. Am. Chem. Soc. 140, 14010–14014 (2018)

    2018

  64. [64]

    & Furukawa, S

    Nakaya, Y., Hirayama, J., Yamazoe, S., Shimizu, K.-I. & Furukawa, S. Single-atom Pt in intermetallics as an ultrastable and selective catalyst for propane dehydrogenation. Nat. Commun. 11, 2838 (2020)

    2020

  65. [65]

    Chang, X. et al. Microenvironment engineering of non-noble metal alloy for selective propane dehydrogenation. Chem 11, 102294 (2025)

    2025

  66. [66]

    Zhang, M. et al. Screening and mechanism exploration of non-noble metal Ni 3M catalysts for propane dehydrogenation: the excellence of synergistic effects. J. Phys. Chem. Lett. 15, 3838–3845 (2024)

    2024

  67. [67]

    Zhou, T. et al. Reaction-induced dynamic evolution of PtIn/SiO 2 catalyst for propane dehydro- genation. Nat. Commun. 16, 5153 (2025)

    2025

  68. [68]

    Yao, Y. et al. Suppressing CO x in oxidative dehydrogenation of propane with dual-atom catalysts. Nat. Commun. 16, 4639 (2025)

    2025

  69. [69]

    Zhou, Y. et al. Peripheral-nitrogen effects on the Ru 1 centre for highly efficient propane dehydrogenation. Nat. Catal. 5, 1145–1156 (2022)

    2022

  70. [70]

    Zhou, H. et al. Cobaltosilicate zeolite beyond platinum catalysts for propane dehydrogenation. Nat. Catal. 8, 357–367 (2025). 38

    2025

  71. [71]

    Zhao, D. et al. In situ formation of ZnO x species for efficient propane dehydrogenation. Nature 599, 234–238 (2021)

    2021

  72. [72]

    Park, J. et al. Pt migration–lockup in zeolite for stable propane dehydrogenation catalyst. Nature 643, 691–698 (2025)

    2025

  73. [73]

    & Zhao, C

    Liu, R., Ma, B., Zhang, L., Tian, J. & Zhao, C. CoO x nano-clusters anchored onto 10-membered ring of Silicalite-1 boost propane dehydrogenation. Nat. Commun. 16, 11476 (2025)

    2025

  74. [74]

    Propane dehydrogenation on Ga 2O3-based catalysts: Contrasting performance with coordination environment and acidity of surface sites

    Castro-Fern´ andez, P.et al. Propane dehydrogenation on Ga 2O3-based catalysts: Contrasting performance with coordination environment and acidity of surface sites. ACS Catal. 11, 907–924 (2021)

    2021

  75. [75]

    Zhang, Y. et al. Control of coordinatively unsaturated Zr sites in ZrO 2 for efficient C–H bond activation. Nat. Commun. 9, 3794 (2018)

    2018

  76. [76]

    Gambo, Y. et al. Catalyst design and tuning for oxidative dehydrogenation of propane – a review. Appl. Catal. A: Gen. 609, 117914 (2021)

    2021

  77. [77]

    Yan, H. et al. Tandem In2O3–Pt/Al2O3 catalyst for coupling of propane dehydrogenation to selective H2 combustion. Science 371, 1257–1260 (2021). 39

    2021