Steric clashes in RNA minimization

I have run into an intermittent problem with RNA_minimizer within the rna_denovo application. I am using identical setups in different directories and shells. I have two calculations currently running without problems. When I try the same calculation in a 3rd directory, I am getting high scores for lk_nonpolar and the resulting pdb structures have steric clashes with overlap side chain atoms. There are no errors reported during the calculation, but the results are not usable.

The calculation is a simple rna hairpin using rna_helix to generate the stem.pdb file:

rna_denovo.macosclangrelease -nstruct 100 -params_file motif.params -fasta motif.fasta -out:file:silent motif.out -include_neighbor_base_stacks -minimize_rna true -s stem.pdb -input_res 6-12 18-24 -cycles 5000 -close_loops -close_loops_after_each_move -filter_lores_base_pairs_early -autofilter -filter_chain_closure_halfway -output_res_num 1-24

Note, I am also having problems with rna_minimize alone. Either the molecules explode or the minimization creates steric clashes between base pairs that are defined in the params file. The rna_minimize application behaves as if it is doing 4D annealing rather than a simple gradient based minimization. Is there anyway to increase the penalty for vdw to prevent steric clashes?

I am using the latest build of Rosetta (2014.35.57232) compiled with clang version 6.0 (600.0.54) in Mac OS X 10.10.0 (currently running 10.10.1). I am preparing to test this in linux next.

I have just tested this in linux (Ubuntu 14.04 kernel 3.13.0-39) with Rosetta compiled under gcc 4.8. I did two runs in different directories with the same setup and both ran fine without the obscenely large lk_nonpolar penalties during the RNA_minimizer step in rna_denovo.

Any suggestions or similar observations are welcome. Thanks. Brian