Convergence issue of ML-NEB with newer ASE

[schumannj]

There seem to be again issues with ML-NEB convergence and the newer ASE 3.20 version.

[Ceasea]

The same problem is also found in a large system with more than 500 atoms.

The following is the ML-NEB process for a large system with 500 atoms.

+--------+------+---------------------+---------------------+---------------------+--------------+--------------+----------+
| Method | Step |        Time         | Pred. barrier (-->) | Pred. barrier (<--) | Max. uncert. | Avg. uncert. |   fmax   |
+--------+------+---------------------+---------------------+------------------+--------------+--------------+----------+
| ML-NEB |     0| 2021-09-13 02:24:44 |             0.000000|             0.000000|      0.000000|      0.000000|  0.299826|
| ML-NEB |     1| 2021-09-13 02:24:44 |             0.000000|             0.000000|      0.000000|      0.000000|  3.601528|
| ML-NEB |     2| 2021-09-13 02:24:44 |             1.640500|             1.641060|      0.017420|      0.014140| 26.143701|
| ML-NEB |     3| 2021-09-13 02:24:44 |            29.339590|            29.340160|      0.066310|      0.030400| 64.284782|
| ML-NEB |     4| 2021-09-13 02:24:44 |            35.294020|            35.294580|      0.014110|      0.012200| 34.493145|
| ML-NEB |     5| 2021-09-13 02:24:44 |            63.151730|            63.152300|      0.011940|      0.011240| 34.493145|
| ML-NEB |     6| 2021-09-13 02:24:44 |            54.541570|            54.542130|      0.011850|      0.011100| 34.493145|
| ML-NEB |     7| 2021-09-13 02:24:44 |            48.122110|            48.122670|      0.011780|      0.010990| 34.493145|
| ML-NEB |     8| 2021-09-13 02:24:44 |            43.274660|            43.275220|      0.011720|      0.010920| 34.493145|
| ML-NEB |     9| 2021-09-13 02:24:44 |            39.493850|            39.494410|      0.011700|      0.010870| 34.493145|
| ML-NEB |    10| 2021-09-13 02:24:44 |            36.460720|            36.461280|      0.011650|      0.010820| 34.493145|
| ML-NEB |    11| 2021-09-13 02:24:44 |            33.970780|            33.971340|      0.011630|      0.010780| 34.493145|
| ML-NEB |    12| 2021-09-13 02:24:44 |            31.887980|            31.888550|      0.011590|      0.010750| 34.493145|
| ML-NEB |    13| 2021-09-13 02:24:44 |            30.118460|            30.119020|      0.011610|      0.010730| 34.493145|
| ML-NEB |    14| 2021-09-13 02:24:44 |            28.595370|            28.595930|      0.011580|      0.010700| 34.493145|
| ML-NEB |    15| 2021-09-13 02:24:44 |            27.269790|            27.270350|      0.011560|      0.010680| 34.493145|
| ML-NEB |    16| 2021-09-13 02:24:44 |            26.105080|            26.105640|      0.011550|      0.010660| 34.493145|
| ML-NEB |    17| 2021-09-13 02:24:44 |            25.073210|            25.073770|      0.011530|      0.010640| 34.493145|
| ML-NEB |    18| 2021-09-13 02:24:44 |            24.152370|            24.152930|      0.011510|      0.010630| 34.493145|
+--------+------+---------------------+---------------------+------------------+--------------+--------------+----------+

However, the general CI-NEB for the same system converges within 77 hours. ML-NEB has difficulty to converge.

In small system with less than 100 atoms, ML-NEB converges fast than general CI-NEB (saveing 2/3 time) and the energy path is almost the same.

Any suggestions for large systems are welcome.

[malik-ust]

Facing the same convergence problem. Looking for some solution especially for the system that contains more than 250 atoms.

[mhangaard]

Thanks @schumannj @Ceasea @malik-ust for posting. How many moving atoms are there? I don't think we tested with a very large number of moving atoms. Could any of you provide a minimal working example?