Photons and electrons
Bonds made with light and current
- photoredox/nickel dual catalysis
- electrosynthesis
- AC-PEC
- asymmetric catalysis
Classical cross-coupling reactions usually require high temperatures and elaborate ligands.We take a different route and drive these reactions with light and electricity. A visible-light photocatalyst, an electrode, or the two together mediate the electron and energy transfer that a nickel catalyst needs, so that the nickel can form a C – C or C – heteroatom bond at room temperature, under conditions mild enough for the late-stage modification of complex molecules. Selectivity can also be designed: the triplet energy of the photocatalyst or the ligand on nickel determines whether a multisubstituted olefin is formed as the E or the Z isomer, and the choice of electrodes determines which of several possible products an enyne gives.
The most recent variable we have learned to control is time. In alternating-current photoelectrocatalysis (AC-PEC) the waveform itself becomes a reaction parameter: periodic reversal of the polarity keeps the metal catalyst cycling between the electrode and the solution instead of plating out, while the frequency, the duty cycle, and the silent time give radical generation and the nickel catalytic cycle separate time windows. This strategy now extends to asymmetric couplings, where the aim is to control which enantiomer is formed as well as which bond.
- Nat. Catal. The triplet energy of the photocatalyst controls olefin geometry in multicomponent coupling
- Nat. Synth. The choice of electrodes determines the products of 1,3-enyne difunctionalization
- Nat. Synth. Asymmetric cross-coupling of alcohols under alternating-current photoelectrocatalysis
- Nat. Commun. Asymmetric-waveform alternating current in enantioselective reductive cross-coupling
Mechanism and descriptors
Why it works, from first principles
- DFT
- catalytic cycles
- selectivity origins
- computed descriptors
A reaction that works is only half of the result; the other half is understanding why it works.We study our own reactions computationally, and the same people run both the calculations and the experiments. With DFT and wavefunction analysis we map complete catalytic cycles, establish what a ligand or an additive actually does, and trace chemo-, regio-, and stereoselectivity to their electronic origins, which matters most when the decisive intermediate is too short-lived to observe.
This is also where the mild couplings of Direction I find their explanation: oxidizing the nickel center by one electron, whether photochemically or electrochemically, turns a C – heteroatom reductive elimination that is prohibitively slow at room temperature into a facile one, and the oxidative addition of an alkyl halide to an excited palladium center proceeds with no barrier at all.
A mechanism has to predict as well as explain: intermediates proposed by calculation are tested in the laboratory, and reactions suggested by calculation are run at the bench. A mapped catalytic cycle also yields numbers: the electronic and steric descriptors that reduce a molecule to a form a model can read. These descriptors are the input to the third direction.
- Angew. Chem. Int. Ed. Ni(III) and excited-state Ni(II) intermediates facilitate reductive eliminationReview
- Angew. Chem. Int. Ed. Oxidative addition of alkyl halides to photoexcited Pd(0)
- J. Am. Chem. Soc. Mechanistic insight into photoredox–nickel–HAT triple catalysis
- Nat. Chem. The olefin slow-release pool mechanism, proposed and validated computationally
AI for Science
A laboratory that improves itself
- Phase I complete
- self-driving laboratory
- dry – wet closed loop
- physical RSI
For us, AI for Science means a laboratory whose models improve through real experiments.The bottleneck in synthesis is the gap between the molecules we can design and the molecules we can actually make. We are building a self-driving laboratory as a dry – wet closed loop. On the wet side, robotic synthesis and analysis stations run and analyze reactions in parallel, returning the results as structured data. On the dry side, models built on the descriptors computed in Direction II weigh those results and, using Bayesian optimization and active learning, select the experiment most likely to correct their current understanding. That experiment is then run at the bench. Closing the loop in this way is what distinguishes a self-driving laboratory from a robot working through a list.
A model that learns only from the literature can remain wrong indefinitely, whereas a model whose proposals are carried out in the physical world and answered by measurement is retrained on every outcome before it proposes the next one. This is what we mean by physical recursive self-improvement (physical RSI). The chemist keeps the part that no machine can take over: deciding which questions are worth putting into the loop.
The loop keeps turning
Read the papers, or join the loop
The papers mentioned above are collected on the Publications page, and current openings for PhD students, postdoctoral fellows, and interns are listed on the Join page.