1: \begin{abstract}

2: A physics-informed neural network (PINN), which has been recently proposed by Raissi \etal [{\em J. Comp. Phys.} 378, pp. 686-707 (2019)], 

3: is applied to the partial differential equation (PDE) of liquid film flows.

4: The PDE considered is the time evolution of the thickness distribution $h(x,t)$ owing to the Laplace pressure,

5: which involves 4th-order spatial derivative and 4th-order nonlinear term.

6: Even for such a PDE, it is confirmed that the PINN can predict the solutions with sufficient accuracy.

7: Nevertheless, some improvements are needed in training convergence and accuracy of the solutions.

8: The precision of floating-point numbers is a critical issue for the present PDE.

9: When the calculation is executed with a single precision floating-point number, 

10: the optimization is terminated due to the loss of significant digits.

11: %

12: Calculation of the automatic differentiation (AD)

13: dominates the computational time required for training, and becomes exponentially longer with increasing order of derivatives.

14: By splitting the original 4th-order one-variable PDE into 2nd-order two-variable PDEs,

15: the computational time for each training iteration is greatly reduced.

16: The sampling density of training data also significantly affects training convergence.

17: For the problem considered in this study, 

18: improved convergence was obtained

19: by allowing the sampling density of training data to be greater in earlier time ranges, 

20: where the rapid diffusion of the thickness occurs.

21: \end{abstract}

22: