Weighted active space protocol for multireference machine-learned potentials.

Area of Science:

• Computational Chemistry
• Quantum Mechanics
• Materials Science

Background:

• Multireference methods accurately describe electronic correlation but are computationally expensive for molecular dynamics.
• Machine-learned interatomic potentials (MLPs) can reduce computational cost but require consistent training data.
• Active space selection in multireference calculations is sensitive to nuclear configurations, hindering consistent data generation.

Purpose of the Study:

• To introduce a systematic approach for consistent active space assignment in multireference calculations across diverse configurations.
• To develop a data-efficient active learning cycle for training MLPs using multireference data.
• To enable accurate and efficient modeling of complex catalytic dynamics.

Main Methods:

• Introduction of the Weighted Active Space Protocol (WASP) for consistent active space selection.
• Integration of WASP with machine-learned interatomic potentials (MLPs) and enhanced sampling techniques.
• Application to TiC+-catalyzed C-H activation of methane, a system with significant multireference character.

Main Results:

• Demonstrated the feasibility of training MLPs on multireference data using the WASP-enhanced active learning cycle.
• Successfully modeled the challenging TiC+-catalyzed C-H activation of methane.
• Achieved accurate and efficient simulation of catalytic dynamics.

Conclusions:

• The WASP protocol overcomes the limitations of active space sensitivity in multireference calculations.
• The proposed framework enables efficient and accurate modeling of complex reactive processes.
• Establishes a new paradigm for simulating systems beyond the reach of conventional electronic-structure methods.