Isayev Lab

Reactions & Reactivity

ML-accelerated reaction modeling enabling exploration of chemical transformations at unprecedented scale

Reactions & Reactivity

Foundation: Quantum Chemistry of Reactions

Our work builds on extensive quantum chemical studies of reaction mechanisms:

  • Energetic materials: decomposition pathways and sensitivity
  • Organic transformations: bond-forming and bond-breaking processes
  • Biological systems: enzyme catalysis and metabolic pathways
  • Transition metal catalysis: organometallic reaction mechanisms

These studies using density functional theory and ab initio molecular dynamics provide the foundation and validation for our machine learning approaches.

ML Potentials for Reactive Chemistry

AIMNet2-rxn Framework

A central contribution involves adapting machine learning interatomic potentials to handle reactive chemistry. The AIMNet2-rxn model, trained on approximately 4.7 million DFT calculations, achieves reaction modeling roughly six orders of magnitude faster than reference quantum mechanical methods while retaining 1-2 kcal/mol accuracy—the precision needed for reliable mechanism prediction.

Key Capabilities

  • Accurate description across bond formation/breaking
  • Transition state location and barrier prediction
  • Intrinsic reaction coordinate following
  • Treatment of radical intermediates and reactive species
  • Coverage of organic reactions and main-group chemistry

Reaction Network Exploration

Batched Nudged Elastic Band Method

We developed methodology enabling systematic exploration of interconnected reaction pathways at unprecedented scale. This batched approach allows exploration of hundreds of reaction pathways in parallel, revealing:

  • Complete reaction networks (not just single pathways)
  • Competing mechanisms and selectivity determinants
  • Side reactions and byproduct formation
  • Kinetically vs. thermodynamically controlled processes

Glucose Pyrolysis Case Study

Demonstration on glucose thermal decomposition revealed complex networks of interconnected transformations, identifying:

  • Major decomposition pathways
  • Key intermediates and branch points
  • Rate-limiting steps
  • Product distribution explanations

Transition State Prediction

Neural network potentials enable rapid prediction of selectivity in complex reactions:

Ring-Forming Reactions

Evaluation of transition state energetics for competing reaction pathways predicts:

  • Regioselectivity (which atoms connect)
  • Stereoselectivity (spatial arrangement of products)
  • Product distributions under different conditions
  • Effects of substituents on reactivity

Applications

  • Reaction design and optimization
  • Catalyst screening
  • Protecting group strategy
  • Synthetic route planning

Catalysis Applications

AIMNet2-Pd Framework

Extension to palladium-catalyzed cross-coupling reactions, maintaining accuracy within 1-2 kcal/mol of quantum mechanical calculations while enabling:

  • Mechanism elucidation for Pd catalysis
  • Ligand effects on reactivity and selectivity
  • High-throughput catalyst screening
  • Prediction of optimal reaction conditions

Broader Impact

Machine learning potentials for catalysis enable:

  • Rational catalyst design
  • Understanding of structure-activity relationships
  • Prediction of novel catalyst performance
  • Acceleration of catalyst discovery

Kinetics & Rate Prediction

Integration with Kinetic Modeling

Combining ML potentials with transition state theory and kinetic modeling enables high-throughput reaction rate prediction:

  • Arrhenius parameters from computed barriers
  • Temperature-dependent rate constants
  • Competition between pathways
  • Kinetic isotope effects

Experimental Kinetics

Recent work on amide coupling reactions uses graph neural networks trained directly on experimental kinetic data, enabling:

  • Prediction of reaction rates from structure alone
  • Identification of slow steps in complex mechanisms
  • Optimization of reaction conditions
  • Understanding of substituent effects on kinetics

Reaction Prediction

Forward Prediction

Given reactants and conditions, predict:

  • Major products and byproducts
  • Reaction mechanisms
  • Optimal conditions (temperature, solvent, catalyst)
  • Potential side reactions

Retrosynthetic Analysis

Working backward from target molecules:

  • Identification of synthetic routes
  • Evaluation of step feasibility
  • Prediction of yields and selectivity
  • Integration with synthesis planning

Applications

Machine learning for reactions enables:

Drug Discovery

  • Metabolic stability prediction
  • Covalent inhibitor design
  • Prodrug activation mechanisms
  • Drug-drug interaction predictions

Process Chemistry

  • Route optimization and scale-up
  • Impurity prediction and control
  • Green chemistry alternatives
  • Continuous flow optimization

Materials Synthesis

  • Polymerization mechanisms
  • Degradation pathways
  • Post-synthetic modifications
  • Stability predictions

Energy Applications

  • Fuel combustion mechanisms
  • Battery electrolyte stability
  • Catalyst deactivation pathways
  • Energy storage reactions

Methodological Advances

Our reaction modeling research emphasizes:

  • Accuracy: Chemical accuracy (1-2 kcal/mol) for reliable predictions
  • Speed: Orders of magnitude faster than quantum chemistry
  • Transferability: Generalizing across reaction types
  • Uncertainty: Knowing when predictions are reliable
  • Integration: Seamless workflows from structure to kinetics

Future Directions

Emerging priorities in reaction modeling:

  • Photochemical reactions and excited states
  • Electrochemistry and electron transfer
  • Enzyme catalysis and protein environments
  • Solvent effects and explicit solvation
  • Machine learning for reaction condition optimization
  • Integration with automated synthesis platforms
  • Autonomous reaction discovery

Software Tools

Key tools for reaction modeling:

  • AIMNet2-rxn: Reactive ML potential
  • AIMNet2-Pd: Palladium catalysis
  • Reaction network tools: Automated pathway exploration
  • Integration with: ASE, RDKit, molecular dynamics packages

Impact

Our methods enable chemists to:

  • Explore reaction mechanisms at unprecedented scale
  • Screen thousands of catalysts computationally
  • Predict reaction outcomes before synthesis
  • Understand selectivity at molecular level
  • Design new reactions rationally

Collaborations

This work involves partnerships with:

  • Synthetic organic chemists
  • Catalysis researchers
  • Process development groups
  • Pharmaceutical companies
  • Software developers and computational scientists

Publications

See our publications page for detailed research on reaction modeling, transition state prediction, catalysis, and kinetics using machine learning methods.