Overview¶
- Inspiration - Anatomy of a PDE
- Data Discretization
- I - State Space Models
- II - Spatial Operators
- III - Reduced Order Models
- IV - TimeSteppers
Anatomy of a PDE¶
- State
- Discretization
- Equation of Motion - Spatial Operators
- Initial Conditions
- Boundary Conditions
- TimeSteppers
Data Discretization¶
- Unstructured, e.g., Weather Stations
- Irregular, e.g., Continents, Countries...
- Curvilinear, e.g., 2D Lat-Lon Coordinates, SWATHs
- Rectilinear, e.g., Geographic Coordinates
- Regular, e.g., Images, GeoStationary Satellites
Parameterization¶
I - State Space Models¶
- Emission Distribution, yt∼p(yt∣zt,θ)
- Transition Distribution, zt∼p(zt∣zt−1,xt,θ)
- Initial Distribution, z0∼p(z0∣θ)
- Parameter Distribution, θ∼p(θ)
II - Spatial Operators¶
- Differentiation
- Exact - Symbolic, AutoDiff
- Approximate - Finite Difference/Volume
- Neural Operators
- Fully Connected
- AutoDifferential + Symbolic
- Convolutions + Finite Difference
- Spectral Convolutions + PseudoSpectral
- Graphical NNs + Finite Element
- Free-Form
- Free-Form Complexity
- Linear
- Basis Function
- Non-Linear
III - Reduced Order Models¶
- Encoder + Decoder, z=f(u),u=g(z)
- Bijections, Surjections, Stochastic
- Separable (POD), u(s,t)=∑ϕ(s)ψ(t)
- Structured
IV - TimeSteppers¶
- Integration
- AutoRegressive
- Taylor - Euler
- Sigma Points - Unscented
- Quadrature - Runge-Kutta
- Monte Carlo
- Hybrid - Leap-Frog, Adam-Bashforth