Table 2.

Averaged MAE, NMAE, and $R^2$ scores for two properties $a_f$ and $b_f$ with respect to network structure in ten-fold cross-validation (2000 epochs).

Network structure	MAE		NMAE, %		$R^2,\%$
Hidden layer $\times$ unit	$a_f$ (kPa)	$b_f$	$a_f$	$b_f$	$a_f$	$b_f$
$1\times 64$	6.991	3.125	7.035	16.101	88.387	49.855
$1\times 128$	7.258	3.083	7.320	15.557	87.377	53.843
$1\times 256$	5.852	3.142	6.069	15.953	88.761	52.704
$1\times 512$	5.553	2.740	5.664	13.863	91.943	63.438
$1\times 1024$	5.444	2.353	5.523	11.857	91.120	67.213
$2\times 64$	5.193	2.627	5.219	13.335	93.749	68.647
$2\times 128$	3.100	2.073	3.152	10.765	97.606	74.895
$2\times 256$	2.499	1.708	2.550	8.645	98.426	84.317
$2\times 512$	2.214	1.281	2.228	6.486	98.395	89.331
$\mathbf{2}\times \mathbf{1024}$	1.755	1.084	1.786	5.558	99.214	91.481
$3\times 64$	4.505	2.225	4.613	11.164	94.505	75.674
$3\times 128$	2.498	1.572	2.614	8.002	97.754	84.299
$3\times 256$	1.958	1.238	1.981	6.350	98.888	89.948
$3\times 512$	2.313	1.324	2.343	6.742	98.388	88.298
$3\times 1024$	2.542	1.767	2.568	9.048	97.764	75.940

The 2 × 1024 structure had the best prediction and was used in the ML model.