Table A.1.

Performance metrics on all assemblies of the finely-discretized dataset, for three of the studied MOGP models (all with q = 16 latent processes) and two baselines. The best and worst value for each metric are highlighted in green and red respectively. Definitions of the metrics are given in Section 3.2. The training time T_train is defined at the level of one assembly. PLMC and VLMC were trained on GPU, with minibatch training in the latter case. The α-CI and PVA metrics are missing for the PLMC are missing because predicted variance was not computed with this model: our current implementation requires the formation of a full p × p noise matrix for this, which causes a memory error in this setting.

Model	R2	RMSE	Errmax	Err${}_{\mathrm{mean}}^{\mathrm{\Sigma }}$	Err${}_{\max }^{\mathrm{\Sigma }}$	α-CI	PVA	Ttrain (s)	ℛ
				(cm−1)	(cm−1)
VLMC	0.185	1.86 ⋅ 10−3	0.189	2.04 ⋅ 10−4	0.543	1.00	−5.90	8, 010	266
PLMC	0.984	2.43 ⋅ 10−3	0.169	1.69 ⋅ 10−4	0.423	–	–	122	266
Lazy-LMC	0.999	2.97 ⋅ 10−3	0.145	1.75 ⋅ 10−4	0.469	1.00	−11.5	5 ⋅ 10−4	266
BPR	0.997	1.19 ⋅ 10−1	0.758	1.78 ⋅ 10−4	0.348	0.987	−1.48	6310	13
Training-less SOGP	0.999	2.86 ⋅ 10−3	0.111	1.24 ⋅ 10−4	0.434	0.999	−6.03	12 200	21