Fig. 2: A 3D-space, 4D-parameter, and time (i.e., 8 inputs) parameterized heat transfer equation is solved with INN to simulate a 10 mm single-track laser powder bed fusion (L-PBF) metal AM. | Nature Communications

Fig. 2: A 3D-space, 4D-parameter, and time (i.e., 8 inputs) parameterized heat transfer equation is solved with INN to simulate a 10 mm single-track laser powder bed fusion (L-PBF) metal AM.

From: Unifying machine learning and interpolation theory via interpolating neural networks

Fig. 2

Detailed problem definition and explanation can be found in the Discussion section. We compare the single-scale finite-difference method (FDM) solver28, the variational multiscale FEM solver12, and the INN solver with CANDECOMP/PARAFAC (CP) decomposition and Q = 2. The data points for the first two methods with dashed marker edges are estimated as they are intractable with a single GPU, while those of INN with a solid marker edge are computed. All benchmarks were conducted with a single GPU, NVIDIA RTX A6000 GPU with 48 GB VRAM.

Back to article page