Chemical Space Exploration with Artificial "Mindless" Molecules.

Area of Science:

• Computational chemistry
• Quantum chemistry
• Materials science

Background:

• Accurate prediction of chemical reaction energies is crucial for computational chemistry.
• Existing molecular benchmarks may not adequately represent chemical space beyond conventional structures.
• Developing robust computational models requires diverse and challenging datasets.

Purpose of the Study:

• Introduce MindlessGen, a novel Python tool for generating chemically diverse molecules.
• Present the MB2061 benchmark set, featuring high-level computational data for H2-promoted decomposition reactions.
• Evaluate the performance of various computational methods, including density functional approximations (DFAs), semiempirical methods, and machine learning potentials.

Main Methods:

• Utilized MindlessGen for random atomic placement and geometry optimization to create novel molecular structures.
• Generated high-level PNO-LCCSD(T)-F12 reference data for 2061 molecules in the MB2061 set.
• Assessed the accuracy of different DFAs, semiempirical methods, and machine learning potentials against the reference data.

Main Results:

• The MB2061 set provides a challenging benchmark, revealing performance trends of computational methods.
• No consistent relationship was found between DFA parametrization strategy and accuracy on this set.
• A Jacob's ladder trend was observed for DFAs, with ωB97X-2 showing the lowest mean absolute error (MAE) of 8.4 kcal·mol⁻¹.
• r²SCAN-3c offered a cost-efficient alternative with an MAE of 19.6 kcal·mol⁻¹.

Conclusions:

• MindlessGen and the MB2061 benchmark set enable rigorous testing of computational chemistry models.
• The study highlights the performance variations of different methods, particularly DFAs, on novel molecular structures.
• Findings guide the selection and development of accurate and efficient computational tools for chemical research.