PINN supervision data

[floatlazer]

Hi,

From the code it seems that PINN samples 30% of ground truth data from the solution domain and use them to supervise training. If I understand correctly, the original PINN does not require supervision data. Is this the desired behavior? I could not locate related descriptions in the paper.

PDEBench/pdebench/models/pinn/train.py

Lines 162 to 181 in cbdeeed

	ratio = int(len(dataset) * 0.3)
	data_split, _ = torch.utils.data.random_split(
	dataset,
	[ratio, len(dataset) - ratio],
	generator=torch.Generator(device="cuda").manual_seed(42),
	)
	data_gt = data_split[:]
	bc_data = dde.icbc.PointSetBC(data_gt[0].cpu(), data_gt[1], component=0)
	data = dde.data.TimePDE(
	geomtime,
	pde_swe2d,
	[bc, ic_h, ic_u, ic_v, bc_data],
	num_domain=1000,
	num_boundary=1000,
	num_initial=5000,
	)

Thanks!

[leiterrl]

Hi,

you are absolutely correct that PINNs don't necessarily need supervision data and can be thought of as a standalone PDE solver. However, we did include some observational data here since the other baseline methods are purely based on supervision and felt that this is a more fair comparison.

[floatlazer]

Thank you for the reply. I may not understand the code completely. Is supervision data sampled from the entire time domain or just from the first few time steps? The first case would look less realistic since it requires knowing the complete PDE solution before hand.

[finger-monkey]

Hi, do you know how to draw pdebench_examples.PNG? This is the visualization picture shown by the author on the code homepage