Chen Zhu Research Group Eastern Institute of Technology, Ningbo
AI for Science automated chemistry
We are building a self-driving laboratory, a dry – wet closed loop in which organic synthesis makes molecules, computation explains them, and AI and automation choose the next experiment. Each proposal is tested at the bench before the model proposes again. For us, AI for Science means physical RSI: recursive self-improvement in the physical world.
The discovery loop
measureexplainchoose
The laboratory works as a single closed loop of making, understanding, and deciding. Automated stations produce real experimental data: yields, selectivities, and spectra measured at the bench rather than simulated. Computation interprets these results in terms of reaction mechanisms and molecular descriptors, an active-learning model uses them to choose the next experiment, and the outcome of that experiment is fed back to the model. Each cycle leaves the model better informed than the one before, so the loop improves with every turn.
The chosen conditions are sent straight back to the bench without manual intervention, and the new results update the model before the next cycle begins.
Three ways into the same question
Our central question is how organic synthesis can become more selective, more sustainable, and more predictable. We approach it through new reactivity, mechanistic understanding, and automated learning, three lines of work that each inform the others.
We develop synthetic methods around two traceless reagents, photons and electrons. Transition-metal catalysis driven by light, by electric current, or by both forms C – C and C – heteroatom bonds under mild conditions, with the selectivity determined by the catalyst, the ligand, and the way the energy is delivered. Alternating-current photoelectrocatalysis extends this control into the time domain.
Computation is how the group understands its own reactions. With DFT and wavefunction analysis we map catalytic cycles, locate transition states, and trace reactivity and selectivity to their electronic origins, both to explain what is observed in the laboratory and to predict what is worth trying next. The same calculations turn molecules into descriptors that machine-learning models can work with, which is how theory enters the automated loop.
The range of possible reaction conditions in modern synthesis is far too large to explore one variable at a time. Our answer is a self-driving laboratory, in which robotic stations run reactions and analyses in parallel, the results are recorded as structured data, and active-learning algorithms choose the experiment from which the model will learn most. This is AI for Science as we practice it: a dry – wet closed loop that changes how we make discoveries.
Latest from the lab
The fourth doctoral cohort is enrolled jointly with Shanghai Jiao Tong University; both students come from Lanzhou University.
Zhiyang Lin joined the group in 2023 as its first postdoctoral fellow and is now a research assistant professor.
Members: Zhiyang Lin awarded an NSFC Young Scientists Fund grant
Congratulations to first author Cai Zhai on a combined mechanistic and experimental study of how ligand class and radical identity govern radical sorting in nickel catalysis.
Read the paper: Cai Zhai's radical-sorting study appears in Inorganic Chemistry
Selected publications
These three papers show the range of the group’s work: controlling a reaction in the time domain, explaining a mechanism, and switching the stereochemical outcome.
Nature Synthesis2025 · 4 · 1534
This study reports the first asymmetric cross-couplings driven by alternating current. Periodic reversal of the polarity prevents the metal catalyst from depositing on the electrode, and chiral C – C and C – heteroatom bonds are formed with up to 99% ee. It marks the start of the group’s work on time-domain control with alternating current.
Nature Chemistry2026 · 18 · 656
Programmable electrochemistry opens strained rings and installs several functional groups in a single operation. DFT modeling of the complete reaction pathway was used to propose the olefin slow-release pool mechanism and to validate it computationally.
Nature Catalysis2019 · 2 · 678
Nickel catalysis combined with single-electron transfer and triplet-energy transfer gives multisubstituted olefins of defined geometry; the triplet energy of the photocatalyst determines whether the E or the Z isomer is formed.
Bring the next question into the loop
We welcome researchers who want to combine rigorous organic chemistry with computation, AI, and laboratory automation, as well as collaborators from any field who can help make this loop more capable.
