Skip to content
DraconiticResearch
Search Sources API
203,234 studies31 sourcesUpdated 20 min ago
PreprintCC BYChemRxiv

Constructive chemical discovery through autonomous reasoning and experiment

  • Yuzhi Richie Xu1,2,3
  • Max Wang4,5
  • Zilong Yan6,7
  • Yu Jiang6,7
  • Yong Lu6,7
  • Yuting Wang6,7
  • Haohui Que1
  • Zhengming Zhang1
  • Letian Chen1
  • Zhenjia Gan6,7
  • Haotian Wu2,3
  • Yilin Wei8
12 more authors
  • Chengyi Xing8
  • Wenxi Zhai1
  • Bowen Li1
  • Daqian Bian9,10
  • Yuli Li11,12
  • Chengwei Ju13,14
  • Jiankai Ge15,16
  • Yingcheng Wu17,18
  • Yuanqing Wang19,20
  • Pheng Ann Heng2,3
  • Dong Xing6,7
  • Tong Zhu1,6,7
20 affiliations
  1. 1Shanghai Innovation Institute
  2. 2Department of Computer Science and Engineering
  3. 3Chinese University of Hong Kong
  4. 4School of Mathematics, Computing, and Data Science
  5. 5New York University
  6. 6Shanghai Engineering Research Center of Molecular Therapeutics and New Drug Development
  7. 7SII–ECNU Joint Autonomous Chemistry Laboratory, School of Chemistry and Molecular Engineering, East China Normal University
  8. 8AINORO Inc.
  9. 9Department of Chemistry
  10. 10University of Chicago
  11. 11Department of Chemistry and Chemical Biology
  12. 12Harvard University
  13. 13Division of Biology and Biological Engineering
  14. 14California Institute of Technology
  15. 15Department of Chemical and Biological Engineering
  16. 16Northwestern University
  17. 17Department of Pathology
  18. 18Stanford University
  19. 19Acceleration Consortium
  20. 20University of Toronto
Open at sourcePDFDOICite

Abstract

Making a new reaction often requires chemical knowledge that does not yet exist. Self-driving laboratories and language-model agents now plan, run and optimize experiments rapidly1–3, but deciding what to ask next still requires a chemist in the loop4. Here we report AIChemist, a robotic chemist for what we call constructive chemical discovery. It decides what to ask next from its own results, settles each question by calculation or experiment, and builds the next design on the answer. Given a silyl-radical-relay concept, AIChemist developed an asymmetric three-component silylcyanation: it established that the solvent effect depended on the oxidant formulation and built the next formulation on this finding, raising the yield from 20% to 67% at 88% enantiomeric excess. For an intermediate of the insecticide broflanilide, it designed a photoredox/nickel C–N coupling route free of nitro-group transformations, at lower estimated cost than the conventional route. When the base favoured by calculation performed poorly, it withdrew that premise and rebuilt the search around the reaction medium, where non-polar toluene raised the product's HPLC yield from 43% to 86%. In both cases, each conclusion extended what the system knew and became the starting point for the next. These results demonstrate a framework for autonomous chemical discovery in which chemical understanding and experimental design evolve together within a closed loop.

Related

  1. Hierarchical Multi-agent Large Language Model Reasoning for Autonomous Heterogeneous Catalyst Discovery
    Samuel Rothfarb, Megan Christina Davis, Ivana Matanović +3
    npj Computational MaterialsPeer-reviewedCC BY-NC-NDFull text14 citedMaterials & polymersOpenAlex
  2. Autonomous in-silico inorganic materials discovery via multi-agent physics-aware scientific reasoning
    Alireza Ghafarollahi, Markus J. Buehler
    npj Computational MaterialsPeer-reviewedCC BY-NC-ND 4.0Full text19 citedMaterials & polymersOpenAlexPMCEurope PMC
  3. Agentic AI for Scientific Reasoning in Autonomous Quantum Sensing Experiments
    Takuya Isogawa, Ryotaro Okabe, Nutdech Phadetsuwannukun +2
    arXiv (Cornell University)PreprintLicence not statedFull text9 citedMaterials & polymersOpenAlex
  4. How Chemical Evidence Shapes Mechanistic Reasoning and Prediction by Large Language Models: Fine-Grained Inference in Nucleoside-Derived Supramolecular Gelation
    Ying Wang, Xi Yu
    ChemRxivPreprintLicence not statedGeneral chemistryChemRxiv
  5. Mechanism-grounded fragmentation trees enable explicit chemical reasoning for electron ionization mass spectrum prediction
    Hongyan Sun, Shuofan Li, Quan Shi +3
    ChemRxivPreprintCC BYFull textPhysical chemistryChemRxiv

Cite

BibTeXRIS
BibTeX
@misc{xu2026constructive,
  title = {{Constructive chemical discovery through autonomous reasoning and experiment}},
  author = {Yuzhi Richie Xu and Max Wang and Zilong Yan and Yu Jiang and Yong Lu and Yuting Wang and Haohui Que and Zhengming Zhang and Letian Chen and Zhenjia Gan and Haotian Wu and Yilin Wei and Chengyi Xing and Wenxi Zhai and Bowen Li and Daqian Bian and Yuli Li and Chengwei Ju and Jiankai Ge and Yingcheng Wu and Yuanqing Wang and Pheng Ann Heng and Dong Xing and Tong Zhu},
  howpublished = {ChemRxiv},
  year = {2026},
  month = oct,
  doi = {10.26434/chemrxiv.15011027/v1},
  url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15011027/v1},
  note = {CC BY},
}
RIS
TY  - UNPB
TI  - Constructive chemical discovery through autonomous reasoning and experiment
AU  - Yuzhi Richie Xu
AU  - Max Wang
AU  - Zilong Yan
AU  - Yu Jiang
AU  - Yong Lu
AU  - Yuting Wang
AU  - Haohui Que
AU  - Zhengming Zhang
AU  - Letian Chen
AU  - Zhenjia Gan
AU  - Haotian Wu
AU  - Yilin Wei
AU  - Chengyi Xing
AU  - Wenxi Zhai
AU  - Bowen Li
AU  - Daqian Bian
AU  - Yuli Li
AU  - Chengwei Ju
AU  - Jiankai Ge
AU  - Yingcheng Wu
AU  - Yuanqing Wang
AU  - Pheng Ann Heng
AU  - Dong Xing
AU  - Tong Zhu
PY  - 2026
DA  - 2026/10/25
PB  - ChemRxiv
DO  - 10.26434/chemrxiv.15011027/v1
UR  - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15011027/v1
AB  - Making a new reaction often requires chemical knowledge that does not yet exist. Self-driving laboratories and language-model agents now plan, run and optimize experiments rapidly1–3, but deciding what to ask next still requires a chemist in the loop4. Here we report AIChemist, a robotic chemist for what we call constructive chemical discovery. It decides what to ask next from its own results, settles each question by calculation or experiment, and builds the next design on the answer. Given a silyl-radical-relay concept, AIChemist developed an asymmetric three-component silylcyanation: it established that the solvent effect depended on the oxidant formulation and built the next formulation on this finding, raising the yield from 20% to 67% at 88% enantiomeric excess. For an intermediate of the insecticide broflanilide, it designed a photoredox/nickel C–N coupling route free of nitro-group transformations, at lower estimated cost than the conventional route. When the base favoured by calculation performed poorly, it withdrew that premise and rebuilt the search around the reaction medium, where non-polar toluene raised the product's HPLC yield from 43% to 86%. In both cases, each conclusion extended what the system knew and became the starting point for the next. These results demonstrate a framework for autonomous chemical discovery in which chemical understanding and experimental design evolve together within a closed loop.
ER  -