You are on page 1of 1

̃ 𝑚,𝑛 + 𝑋𝑚,𝑛 .

𝑦𝑚,𝑛 = [IDFT(𝑍)] (5)

The complete mathematical description of the AFNO model is beyond the scope of this material,
and we refer you to Modulus User Documentation for additional details.

1.3 Physics-informed Neural Operators


The Physics-Informed Neural Operator (PINO) was introduced by Li et al. in this paper. The
PINO approach for surrogate modeling of PDE systems effectively combines the data-informed
supervised learning framework of the FNO with the physics-informed learning framework. The
PINO incorporates a PDE loss ℒ𝑝𝑑𝑒 to the Fourier Neural Operator. This reduces the amount of
data required to train a surrogate model, since the PDE loss constrains the solution space. The
PDE loss also enforces physical constraints on the solution computed by a surrogate ML model,
making it an attractive option as a verifiable, accurate and interpretable ML surrogate modeling
tool.
To explain the key concepts behind PINO, consider a stationary PDE system. Following the
notation used in the paper, we consider a PDE represented by,

𝒫(𝑢, 𝑎) = 0, in 𝐷 ⊂ ℝ𝑑 , (6)

𝑢 = 𝑔, in 𝜕𝐷. (7)

Here, 𝒫 is a Partial Differential Operator, 𝑎 are the coefficients/parameters and 𝑢 is the PDE
solution.
In the FNO framework, the surrogate ML model is given by the solution operator 𝒢†𝜃 , which maps
any given coefficient in the coefficient space 𝑎 to the solution 𝑢. The FNO is trained in a supervised
fashion using training data in the form of input/output pairs {𝑎𝑗 , 𝑢𝑗 }𝑁
𝑗=1 . The training loss for the
FNO is given by summing the data loss, ℒ𝑑𝑎𝑡𝑎 (𝒢𝜃 ) = ‖𝑢−𝒢𝜃 (𝑎)‖2 over all training pairs {𝑎𝑖 , 𝑢𝑖 , }𝑁
𝑖=1 .

In the PINO framework, the solution operator is optimized with an additional PDE loss given
by ℒ𝑝𝑑𝑒 (𝑎, 𝒢𝜃 (𝑎)) computed over i.i.d. samples 𝑎𝑗 from an appropriate supported distribution in
parameter/coefficient space.
In general, the PDE loss involves computing the PDE operator which in turn involves computing
the partial derivatives of the Fourier Neural Operator ansatz. In general, this is nontrivial. The
key set of innovations in the PINO are the various ways to compute the partial derivatives of the
operator ansatz, viz:
1. Numerical differentiation using a Finite-Difference Method (FDM).
2. Numerical differentiation computed via spectral derivative.
3. Hybrid differentiation based on a combination of first-order “exact” derivatives and second-
order FDM derivatives.
For more details, please refer the Modulus User Documentation.
Now, with this theoretical background, let’s solve the Darcy flow problem using these three ap-
proaches.

You might also like