maruthinh/PiNNPotentialFoam

pinnPotentialFoam solution and test case

Hackers squad of team 2: Jun Liu (liu@mma.tu-darmstadt.de), Guoliang Chai (guoliangchai@stu.xjtu.edu.cn), Maruthi N H (maruthinh@gmail.com)

Our solution for the physics-based-dl part of OpenFOAM Machine Learning Hackathon. We adapted the Neural Network (NN) $\Psi(x,\ y,\ z,\ \theta)$ to map the cell center field $\mathbf{x}= (x,\ y,\ z)$ to the output vector field $\mathbf{o}= (\phi,\ u_x,\ u_y,\ u_z)$, where $\phi$ is the velocity potential and $\mathbf{u}=(\ u_x,\ u_y,\ u_z)$ is the velocity. Different losses are tested:

$$LOSS_{data}=\frac{1}{N_{t}}\sum_{p=1}^{N_t}(||\phi_{p}^{nn}-\phi_{p}||+||u_{x,p}^{nn}-u_{x,p}||+||u_{y,p}^{nn}-u_{y,p}||+||u_{z,p}^{nn}-u_{z,p}||)^2$$

$$LOSS_{residual1}=\frac{1}{N_{t}}\sum_{p=1}^{N_t}(||\nabla \cdot \nabla \phi_p^{nn} - \nabla \cdot \mathbf{u}_p^{nn}||)^2$$

$$LOSS_{residual2}=\frac{1}{N_{t}}\sum_{p=1}^{N_t}(||\nabla \cdot \nabla \phi_p^{nn} - \text{fvc::laplacian(phi)}||)^2+\frac{1}{N_{t}}\sum_{p=1}^{N_t}(||\nabla \cdot \mathbf{u}_p^{nn}- \text{fvc::grad(u)}||)^2 $$

To invesgite the impact of $LOSS$ on the results, three combinations are implemented: $LOSS = LOSS_{data}$, $LOSS = LOSS_{data} + LOSS_{residual1}$, $LOSS = LOSS_{data} + LOSS_{residual1} + LOSS_{residual2}$.

The training data are from the cylinder case of potentialFoam.