Reactions & Reactivity

Understanding and predicting chemical reactivity represents one of the central challenges in computational chemistry. Reactions involve bond breaking and formation, electronic reorganization, and navigation of complex potential energy landscapes—regimes where classical force fields fundamentally fail and quantum chemistry becomes computationally prohibitive for extensive exploration. My research develops machine learning approaches that enable reaction modeling at unprecedented scale while retaining the accuracy required for mechanistic insight and reliable prediction.

Foundations in Reaction Mechanism Studies

My research program has deep roots in computational studies of chemical reactions. Early work applied quantum chemical methods—particularly density functional theory and ab initio molecular dynamics—to investigate reaction mechanisms in diverse systems including thermal decomposition pathways in energetic materials, regioselectivity in organic transformations, redox reactions involving transition metal complexes, and tautomerization and proton transfer in biological contexts.

These studies established both methodological expertise and physical intuition about reaction energy landscapes. Understanding transition state theory, reaction coordinates, and the factors that control selectivity provides essential context for developing machine learning approaches to reactivity prediction. The limitations encountered in this work—particularly the computational cost of exploring complex reaction networks—motivated the development of ML-accelerated reaction modeling.

ML Potentials for Reactive Chemistry

A major theme of my research is adapting machine learning interatomic potentials for reactive chemistry. Standard ML potentials trained on equilibrium configurations often fail catastrophically when bonds break or form, producing unphysical energies and forces. Addressing this challenge requires training data that samples reactive configurations and transition state regions; neural network architectures that remain well-behaved across bonding topology changes; validation protocols that specifically assess accuracy along reaction coordinates; and careful treatment of electronic effects including spin state changes and charge redistribution.

The AIMNet2-rxn model represents a breakthrough in generalized reaction modeling. Trained on approximately 4.7 million range-separated DFT calculations spanning diverse reaction types, AIMNet2-rxn enables reaction modeling roughly six orders of magnitude faster than reference quantum mechanical methods while retaining 1-2 kcal/mol accuracy across reaction coordinates. This combination of speed and accuracy enables exploration at scales that would be impossible with conventional computational chemistry.

Reaction Network Exploration

Beyond modeling individual reactions, my research enables systematic exploration of reaction networks—the interconnected web of possible transformations that molecules can undergo. This capability is essential for understanding complex reaction mixtures and predicting product distributions, identifying unexpected reaction pathways and byproducts, designing synthetic routes and optimizing reaction conditions, and studying degradation, metabolism, and environmental fate.

To exploit GPU parallelism and ML potential efficiency, we developed a batched nudged elastic band (BNEB) method that achieves minimum energy pathway searches at the scale of millions of reactions. This enables deep reaction network analysis that was previously computationally prohibitive. Demonstration on glucose pyrolysis—evaluating an 11-step pathway producing hydroxymethylfurfural, the experimentally observed major product—illustrates the capability to characterize complex multi-step transformations with quantitative energetics.

Transition State Prediction

Transition states represent the critical points that determine reaction rates and selectivity. Accurate transition state prediction requires both locating these stationary points on the potential energy surface and computing their energies with sufficient accuracy to enable meaningful kinetic predictions. Machine learning approaches address both challenges.

Recent work demonstrates that neural network potentials can anticipate the selectivity of intramolecular cyclization reactions—predicting which of multiple possible ring-forming pathways will dominate based on transition state energetics. The ability to rapidly screen many possible pathways, combined with near-quantum accuracy for relative energies, enables computational prediction of selectivity that would be prohibitively expensive with conventional methods.

Catalysis and Organometallic Chemistry

Chemical catalysis involves complex reactivity at metal centers, where electronic structure effects are particularly important and conventional force fields are essentially useless. My research extends ML potential methodology to catalytic systems, including palladium-catalyzed cross-coupling reactions, where AIMNet2-Pd enables rapid computational screening of substrate-catalyst combinations. This framework replaces expensive electronic structure calculations with neural network predictions while maintaining accuracy within 1-2 kcal/mol and structural accuracy of approximately 0.1 Angstrom compared to reference QM calculations. Importantly, transferability extends beyond the monophosphine ligands in the training set to chemically diverse Pd complexes, demonstrating generalization capability essential for practical catalyst screening.

Reaction Kinetics and Rate Prediction

Moving from thermodynamics to kinetics requires accurate activation energies combined with appropriate rate theories. My research develops methods for high-throughput reaction rate prediction that integrate ML potentials with kinetic modeling frameworks.

Recent work on amide coupling reactions demonstrates automated measurement of reaction rates using machine vision and indicator dyes—a technique called PRISM that achieves precision comparable to NMR while measuring twelve rate constants concurrently across more than four orders of magnitude. Computational investigation revealed concerted asynchronous SN2 mechanisms, with base-catalyzed pathways showing the lowest barriers. A graph neural network trained on this dataset achieves high accuracy for out-of-distribution reactants, demonstrating that chemistry-informed representations enable meaningful generalization.

Applications in Synthesis and Discovery

The ability to model reactions rapidly and accurately has direct applications in synthesis planning and chemical discovery. Integration of reaction modeling with molecular design enables assessment of synthetic accessibility for generated compounds, prediction of reaction outcomes before experimental attempts, identification of conditions likely to favor desired products, and understanding of competing pathways that might reduce yield or selectivity.

This reaction-aware perspective increasingly informs my work in drug discovery and materials design, where synthetic feasibility is a critical practical constraint.

Outlook

The future of computational reaction chemistry lies in increasingly autonomous systems that can explore reaction space, identify interesting transformations, and guide experimental validation. Key directions include extending ML potentials to broader catalyst classes including earth-abundant metals; integrating reaction prediction with retrosynthesis and route planning; developing uncertainty-aware methods that know when predictions require experimental validation; and building toward autonomous reaction discovery platforms that iterate between computation and experiment.

The long-term goal is to make computational reaction prediction a routine tool for synthetic chemists—not replacing experimental intuition, but augmenting it with quantitative guidance that enables more efficient exploration of chemical reactivity.