How does the deepxde package solve a pde without training data?

[amiralizadeh1]

I was looking into the examples provided, specially Allen-Cahn forward problem, I realized that the 'data' object contains only 'pde', 'geometry', and number of points on the boundary and domain. Then a model is built according to 'data' and 'net' objects.

My question is, how does DeepXDE solve the PDE without being provided with any training data?

Of course, there is a 'gen_testdata' function but this is only used for testing the model, not training it.

[lululxvi]

See https://doi.org/10.1137/19M1274067