©2001 by P. Blaha, K. Schwarz and J. Luitz
Lstart is a fully relativistic atomic code and thus provides a better (and thus lower) total energy than the LAPW calculations, which have only a scalar relativistic basis (even when SO is added a posteriori).
Starting for second-row elements (this depends a little on the required accuracy) you must calculate the atomic energy not with LSTART, but with LAPW. Create a FCC box with lattice constants of about 30 bohr and your atom at (0,0,0). Use identical RMT, RKMAX, but only 1 k-point and spinpolarization (for non-closed shell atoms).
Since for such a supercell EF will be most likely at negative energies (eg. -0.3 or similar), you may have to adjust the energy parameters in case.in1. Once the scf-cycle has finished, you may need to readjust these energy parameters by hand (compare eigenvalues, fermi energy and :EP labels from your scf file, eventually try the "-in1new 5" switch).
Some atoms (with partly occupied orbitals or in particular transition metal atoms, where the occupation flips between a 4s and 3d states) may be difficult to converge. Use a broadening method (e.g. TEMP 0.005) to get equal partial occupations and thus a stable solution.