(361e) Quantumpioneer: Self-Evolving Machine for High-Throughput Automated Potential Energy Surface Exploration and Closed-Loop Chemical Reactivity Discovery

Acquiring accurate 3D geometries of reactive chemical species and transition states (TS) along reaction pathways over the potential energy surface is a critical prerequisite for computing high-quality thermodynamic and kinetic parameters via quantum chemistry and statistical mechanics. These calculated thermo-kinetic parameters are essential for creating impactful predictive models that can assist pharmaceutical design, high-performance material development, and renewable energy research. However, the search for transition states has historically been a trial-and-error process that demands considerable manual input. Although automated TS search has a rich history of algorithmic development, existing methods like single-ended and double-ended approaches are often computationally demanding, have high failure rates, and are highly localized to specific reaction systems, making it challenging to learn from previous examples and enhance performance. Therefore, efficiently automating the TS search is essential for enabling high-throughput chemical reactivity discovery and streamlining the generation of extensive theoretical reaction datasets that are valuable for developing machine-learning-based chemistry surrogate models. In this presentation, we introduce QuantumPioneer, a self-evolving machine capable of automatically generating validated 3D electronic structures for a given reaction using atom-mapped SMILES as input. QuantumPioneer can enhance its performance over time by learning from past experiences and can be seamlessly integrated into an active learning loop for closed-loop chemical discovery. Moreover, QuantumPioneer can be tailored to specific chemical-exploration tasks and generate extensive theoretical reaction datasets that lay the foundation for future studies with wide-ranging applications. To showcase its capabilities, we employed QuantumPioneer to generate TS at ωB97XD/def2svp level of theory for over 100,000 hydrogen atom abstraction (HAbs) reactions between peroxyl radicals and a diverse set of organic molecules chosen based on their functional groups. This dataset represents one of the largest DFT TS collections for radical-based reactions to date, laying the foundation for future studies with wide-ranging applications. We further refined the dataset with COSMO-RS and DLPNO-CCSD(T1)-F12D single-point calculations and trained D-MPNN models to predict HAbs reaction barriers. We will demonstrate how the machine learning model for HAbs barrier height prediction can potentially be applied to conduct high-throughput virtual screening of oxidative stability of organic molecules and aid better design of high-value chemicals such as active pharmaceutical ingredients and carbon capture amines.