Benchmarking Isomerization Energies for C 5 $$ {\mathrm{C}}_5 $$ - C 7 $$ {\mathrm{C}}_7 $$ Hydrocarbons: The ISOC7...

Highly accurate benchmark databases are critical for the development of robust and computationally efficient electronic structure methods. We introduce the ISOC7 database, a diverse collection of 1308 unique constitutional isomers of ${\text{C}}_5$ - ${\text{C}}_7$ saturated and unsaturated hydrocarbons with reference isomerization energies at the CCSD(T)/CBS level of theory, obtained via the W1-F12 thermochemical protocol. The isomerization energies in this dataset span over 146  $\text{kcal}{\text{mol}}^{-1}$ . This database was used to conduct a rigorous benchmark assessment of a wide hierarchy of computational methods. The performance of 40 contemporary density functional theory (DFT) functionals reveals a general, albeit not monotonic, improvement along the rungs of Jacob's Ladder, with lower-rung GGA and MGGA functionals providing generally poor performance. The range-separated hybrid-meta-GGA functional $\omega$ B97M-D4 emerges as the top DFT performer with a root-mean-square deviation (RMSD) of 1.62  $\text{kcal}{\text{mol}}^{-1}$ . We also evaluated computationally economical semiempirical and tight-binding methods. While traditional semiempirical approaches are inadequate, the modern g-xTB tight-binding method achieves a respectable RMSD of 4.14  $\text{kcal}{\text{mol}}^{-1}$ . Remarkably, the machine-learned neural network potential AIMNet2 delivers exceptional accuracy, achieving an RMSD of 1.67  $\text{kcal}{\text{mol}}^{-1}$ , rivaling the performance of the best DFT functional at a fraction of the computational cost. The ISOC7 database provides a challenging benchmark for advancing the development and validation of both quantum chemical and machine learning methods.