. 2024 Jan 26;128(5):945–957. doi: 10.1021/acs.jpca.3c07129

Table 1. Average MAE, RMSE, and Dataset Size with Standard Deviations from 5-Fold Cross-Validation Given in Parentheses of Combinations of Active Learning Hyperparameters Evaluated by Training cNNPs with an Oxalic Acid CSP Landscape^a.

entry	training quantity	committee size	% uncertain target	batch size	energy MAE (kJ mol^–¹)	energy RMSE (kJ mol^–¹)	final training set size
1	E	6	10.0	30	1.11 (0.05)	1.74 (0.18)	852 (24)
2	E, F	6	10.0	30	1.08 (0.10)	1.50 (0.13)	546 (40)
3	ΔE	6	10.0	30	1.09 (0.11)	1.47 (0.18)	205 (53)
4	ΔE, ΔF	6	10.0	30	0.92 (0.06)	1.20 (0.07)	216 (22)
5	ΔE	6	2.5	30	0.89 (0.11)	1.21 (0.20)	354 (72)
6	ΔE	6	5.0	30	0.97 (0.06)	1.29 (0.10)	252 (45)
7	ΔE	2	5.0	30	1.19 (0.18)	1.61 (0.26)	168 (65)
8	ΔE	18	5.0	30	0.90 (0.06)	1.20 (0.12)	288 (41)
9	ΔE	6	5.0	15	1.01 (0.05)	1.37 (0.17)	216 (15)
10	ΔE	6	5.0	60	0.90 (0.08)	1.38 (0.43)	320 (45)

The cNNPs were trained either on total energy (E)/forces (F) or the difference between the force field (FIT + DMA) values and the reference values (ΔE/ΔF), i.e., Δ-learning. All entries used an uncertainty cutoff of 1.0 kJ mol^–1, with candidates selected by highest uncertainty.

Table 1. Average MAE, RMSE, and Dataset Size with Standard Deviations from 5-Fold Cross-Validation Given in Parentheses of Combinations of Active Learning Hyperparameters Evaluated by Training cNNPs with an Oxalic Acid CSP Landscapea.

Table 1. Average MAE, RMSE, and Dataset Size with Standard Deviations from 5-Fold Cross-Validation Given in Parentheses of Combinations of Active Learning Hyperparameters Evaluated by Training cNNPs with an Oxalic Acid CSP Landscape^a.