Reliability of computed molecular structures

When the structures of 1342 molecules are optimized by 30 methods and 7 basis sets, there appear 289 (21.54%) problematic molecules and 112 (8.35%) failed ones. When 278 problematic molecules are compared, the best methods are BHandH and LC‐wPBE, while B97D, BP86, HFS, VSXC, and HCTH are very unreliable. When 179 problematic molecules are computed with larger basis sets, the smallest mean absolute deviation (MAD) of bond angle (2.3°) is shown by QCISD(T)/cc‐pVTZ, while the smallest MAD of bond length (0.021 Å), the best SUM1 (4.9 unit), and the best SUM2 (2.4 unit) are shown by DSDPBEP86(Full), DSDPBEP86, PBE1PBE‐D3, MP2, and MP2(Full) in combination with aug‐cc‐pVQZ, cc‐pVQZ, Def2QZVP, Def2TZVPP, and/or 6–311++G(3df,3pd). Very large basis sets, for example, larger than cc‐pVTZ usually have to be used to obtain very good structures and the performances of many density‐functional theory methods are encouraging. The best results may be the limit of modern computational chemistry. Although the deviation ΔR2 of HF calculating ClOOCl is close to that of APFD with their corresponding basis sets, its SUM1 and SUM2 (i.e., 7.0 unit and 3.6 unit) are respectively much larger than those (5.0 unit and 2.5 unit) of APFD or CCSD(T) which even exceed those (4.9 unit and 2.4 unit) of DSDPBEP86(Full) calculating the 179 molecules.