. 2023 Jun 8;19(12):3509–3525. doi: 10.1021/acs.jctc.3c00279

Table 5. RMSE Values of Individual HDNNPs and the Ensemble Trained on Reference Data Set C Using Learning on a Stationary Batch of All Training Data and Lifelong Learning^a.

Individual HDNNPs	Stationary data	Lifelong learning
RMSE(E^train)/meV atom^–1	6.1 ± 0.2	6.9 ± 0.6
RMSE(E^test)/meV atom^–1	6.8 ± 0.2	7.8 ± 0.5
RMSE(F_α,n^train)/meV Å^–1	168 ± 3	185 ± 12
RMSE(F_α,n^test)/meV Å^–1	182 ± 4	205 ± 10

Ensemble
RMSE(E)/meV atom^–1	4.3	4.5
RMSE(F_α,n)/meV Å^–1	121	122

In lifelong learning, only the conformations of reference data set B were trained for the initial 1250 epochs and then the additional conformations of reference data set C were added. The CoRe optimizer with β₁^b = 0.7 and lifelong adaptive data selection were applied for 2500 epochs.

Table 5. RMSE Values of Individual HDNNPs and the Ensemble Trained on Reference Data Set C Using Learning on a Stationary Batch of All Training Data and Lifelong Learninga.

Table 5. RMSE Values of Individual HDNNPs and the Ensemble Trained on Reference Data Set C Using Learning on a Stationary Batch of All Training Data and Lifelong Learning^a.