Revisiting medium-sized silicon clusters: new lowest-energy structures for Si₁₈, Si₂₀, Si₃₀ and Si₃₂ and correlated correction of DFT misassignments

Silicon clusters are prototype building blocks of nanoscale silicon materials, linking molecular chemistry with bulk semiconductor properties that underpin optoelectronics and quantum technologies. Herein, we revisit medium-sized silicon clusters through a workflow that combines machine-learned interatomic potentials (MLIPs), genetic-algorithm searches, electronic-structure geometry optimisations, and DLPNO-CCSD(T) benchmarking. The MLIPs were trained on TPSSh/def2-TZVP and r² SCAN-D3/def2-TZVP energies and forces, enabling rapid screening of candidate structures and broad exploration of the configurational landscape. Candidates within 15 kcal/mol of the lowest-energy structure identified by the genetic algorithm were reoptimised at the PBE0/def2-TZVP level, and we computed single-point energies with DLPNO-CCSD(T) in combination with the cc-pVDZ and cc-pVTZ basis sets. This workflow identifies new lowest-energy structures to date for Si18, Si20, Si30, and Si32. Across these clusters, the cc-pVDZ to cc-pVTZ shift in the relative energy difference between the new and previously reported minima ranges from 0.1 to 5.4 kcal/mol. For example, the corresponding gaps for Si20 are 14.7 and 11.5 kcal/mol with cc-pVDZ and cc-pVTZ, respectively, without changing the energetic ordering of the isomers. In contrast, putative new minima identified by DFT for Si17, Si22, Si23, and Si29 were not confirmed by correlated calculations, which placed these candidate structures up to 20 kcal/mol above the corresponding literature minima. This recurring discrepancy reveals a functional sensitivity in which compact cage-like motifs are overstabilised relative to more extended C_s and C_2v and structures. We also report IR fingerprints for the identified low-energy structures to aid experimental assignment. Overall, these results show how MLIP-accelerated exploration, combined with selective high-level validation, enables efficient discovery and reliable assignment of low-energy structures in silicon clusters.