This benchmark tests the performance of algorithms for finding the nearest transition state. The starting point are structures taken from the highest energy image in a DNEB run. Idealy they should all be reasonably close to a transition state (Although there are a few structures which converge to states with positive lowest eigenvalue. Should we delete these from the starting structure database?) A tar file containing the random and minimized structures is (will be) here lj38.tar.gz.
The benchmark requires that each provided Lennard-Jones starting structures be optimized until the norm of the gradient is less that 0.001. Runs that don't reach the required tolerance or those with positive zero eigenvalue are considered failures. If a run was successful, the number of forcecalls necessary to reach the global minimum the first time is recorded.
Avg FCs: Average number of force calls needed to find global minimum (average over successful runs)
min(FCs): Minimum number of force calls needed to find global minimum (best successful run)
max(FCs): Maximum number of forcecalls needed to find the global minimum(worst successful run)
Algorithm | <FCs> | min(FCs) | median(FCs) | max(FCs) | nfailed |
---|---|---|---|---|---|
sella | 70 | 24 | 64 | 158 | 0 |
optim_ffd | 145 | 57 | 119 | 565 | 0 |
pele_ffd | 192 | 59 | 150 | 1488 | 0 |
eon-lanczos-improved1 | 237 | 65 | 181 | 1898 | 0 |
eon-lanczos1 | 321 | 83 | 220 | 8857 | 0 |
eon-dimer1 | 528 | 92 | 388 | 3581 | 0 |
Show/Hide Additional Entries... |
Algorithm | <FCs> | min(FCs) | median(FCs) | max(FCs) | nfailed |
---|---|---|---|---|---|
sella | 53 | 31 | 51 | 108 | 0 |
optim_ffd | 71 | 43 | 70 | 143 | 0 |
pele | 88 | 52 | 78 | 198 | 0 |
optim_cd | 90 | 52 | 86 | 199 | 0 |
eon-lanczos-improved1 | 106 | 71 | 100 | 163 | 0 |
eon-dimer1 | 116 | 83 | 114 | 160 | 0 |
eon-lanczos1 | 143 | 106 | 139 | 292 | 0 |
Show/Hide Additional Entries... |