Research

Research in the Keith Lab centers on a single question: how does electronic structure govern chemical reactivity, and how can we predict it accurately without prohibitive computational cost? 

The group develops and applies quantum chemistry methods to study chemical bonding, reaction mechanisms, molecular adsorption, solvated ion chemistry, and acid/base reactivity, with a focus on the difficult cases where conventional fast methods break down, such as bond dissociation in solution, transition-metal chemistry, and reactive intermediates.

A defining thread of our work is the design of physics-informed, minimally parameterized models that stay chemically interpretable. The group's "bond energy from bond orders and populations" (BEBOP) framework, for instance, recovers density-functional-level accuracy at a fraction of the cost while decomposing molecular energies into physically meaningful bond contributions, with extensions to zero-point energies and vibrational partition functions. Complementary efforts span quantum alchemy for rapid catalyst screening across chemical space, reactive and machine-learned force fields, and the broader integration of machine learning with computational chemistry. Applied together, these tools let the group map reaction pathways and thermodynamics under realistic operating conditions to guide the design of next-generation catalysts for sustainable fuels and chemicals, including carbon dioxide reduction, water oxidation for ozone for disinfection, and reactions scaffolded by earth-abundant metal catalysts.

For recent work we are actively pursuing, see:

• John A. Keith, Valentin Vassilev-Galindo, Bingqing Cheng, Stefan Chmiela, Michael Gastegger, Klaus-Robert Müller, and Alexandre Tkatchenko. "Combining machine learning and computational chemistry for predictive insights into chemical systems." Chem. Rev., 121:9816–9872, 2021. link
  
  This is a broad tutorial review mapping how machine learning and quantum chemistry can be combined for predictive modeling across molecular simulation, catalysis, retrosynthesis, and drug design. 

• Abdulrahman Y. Zamani, Barbaro Zulueta, Andrew M. Ricciuti, John A. Keith, and Kevin Carter-Fenk. "Kohn-Sham density encoding rescues coupled cluster theory for strongly correlated molecules." link

This shows that using a Kohn-Sham reference instead of Hartree-Fock dramatically improves coupled cluster theory for strongly correlated systems, and traces that gain to the one-particle density matrix rather than the orbitals themselves. The resulting KS-CCSD(T) reaches near-chemical accuracy for transition-metal dimers and even recovers the notoriously difficult Cr2 potential energy surface, with a new density-difference diagnostic to flag multireference character at mean-field cost. 

• Remsha Rafiq, Barbaro Zulueta, Hannah Zucco, Ramakrishna Suresh, Jason E. Shoemaker, Michael Call, Daylan Sheppard, Glenn Cormack, John A. Keith, and Götz Veser. "Physics-Informed Descriptors Enable Machine Learning in Data-Sparse Chemical Systems." ChemRxiv (preprint). link

This shows that descriptors built from the group's BEBOP method inject electronic-structure and bonding information into machine learning, enabling accurate predictions from very small datasets. Trained on just 19 experimental compounds, a BEBOP-informed LASSO model predicts capped-diisocyanate deblocking temperatures to about 11 °C RMSE across a 227–323 °C range, where conventional molecular descriptors produce no meaningful correlation, cutting data requirements by an order of magnitude for molecular design problems.

• Barbaro Zulueta, Sonia V. Tulyani, Phillip R. Westmoreland, Michael J. Frisch, E. James Petersson, George A. Petersson, and John A. Keith. "A bond-energy/bond-order and populations relationship." J. Chem. Theory Comput., 18(8):4774–4794, 2022. link

This introduces the BEBOP model, which uses Hartree-Fock bond orders and populations to predict molecular atomization energies with hybrid-DFT-level accuracy at a small fraction of the cost, while decomposing energies into physically interpretable bond contributions. 

• Barbaro Zulueta, Colin D. Rude, Jesse A. Mangiardi, George A. Petersson, and John A. Keith. "Zero-point energies from bond orders and populations relationships." J. Chem. Phys., 162(8), 2025. link

This extends the BEBOP framework to predict molecular zero-point vibrational energies directly from bond orders and populations, bypassing expensive Hessian calculations. 

• Barbaro Zulueta and John A. Keith. "Vibrational partition functions from bond order and populations relationships." ChemPhysChem, 26(14):e202500085, 2025. link 

This Further extends BEBOP to estimate vibrational partition functions, and thus thermochemical corrections, from orbital populations without costly frequency calculations.