[Paper Review] Quantum based machine learning of competing chemical reaction profiles

Stefan Heinen, Guido Falk von Rudorff+1 more|arXiv (Cornell University)|Sep 28, 2020

Machine Learning in Materials Science4 citations

TL;DR

This paper proposes a quantum machine learning framework to predict competing E2 and SN2 reaction outcomes and transition states across chemical compound space. By training on quantum-chemical data, the model reveals Hammond's postulate holds for SN2 but not E2, enabling predictive decision trees for rational reaction design.

ABSTRACT

Kinetic and thermodynamic effects govern the outcome of competing chemical reactions, and are key in organic synthesis. They are crucially influenced, if not dominated, by the chemical composition of the reactants. For two competing exemplary reactions, E2 and SN2, we show how to use quantum machine learning in chemical compound space to rapidly predict outcome and respective transition states for any new reactant. Machine learning model based predictions of reactions in the chemical compound space of reactant candidates affords numerical results suggesting that Hammond's postulate is valid for SN2, but not to E2. The predictions are demonstrated to enable the construction of decision trees for rational prospective experimental design efforts.

Motivation & Objective

• To predict reaction outcomes and transition states for competing E2 and SN2 reactions across diverse reactant compounds.
• To assess the validity of Hammond's postulate in the context of SN2 versus E2 mechanisms using machine learning.
• To develop predictive models that support prospective experimental design in organic synthesis.
• To map reaction profiles in chemical compound space using quantum-chemical data and machine learning.

Proposed method

• The method employs quantum machine learning models trained on quantum-chemical calculations of reactants and transition states.
• It explores the chemical compound space of reactant candidates to generalize predictions across new, unseen reactants.
• The model uses numerical predictions to evaluate whether Hammond's postulate applies to SN2 and E2 mechanisms.
• Decision trees are constructed from model predictions to guide experimental design in organic synthesis.
• The approach integrates quantum chemical data with machine learning to predict kinetic and thermodynamic preferences.

Experimental results

Research questions

• RQ1 Does Hammond's postulate hold for SN2 reactions when predicted by quantum machine learning?
• RQ2 How accurately can machine learning predict transition states and reaction outcomes for competing E2 and SN2 pathways?
• RQ3 To what extent do quantum machine learning models reveal differences in kinetic and thermodynamic control between E2 and SN2 mechanisms?
• RQ4 Can predictive models derived from quantum data enable rational, prospective experimental design in organic synthesis?

Key findings

• Hammond's postulate is validated for SN2 reactions based on machine learning predictions.
• The postulate does not hold for E2 reactions according to the model's predictions.
• The model successfully predicts reaction outcomes and transition states across diverse reactant compounds in chemical space.
• Predictive results enable the construction of decision trees for guiding experimental reaction design.
• Quantum machine learning effectively captures kinetic and thermodynamic effects in competing reaction pathways.