Agnostix / Platform / Train
M-02Agnostix Train

Train
Physics-Aware
Models.

The computational core. Solver-agnostic by design — PINNs, DeepONet, FNO or hybrid classical-plus-learned. The engine is chosen to fit the physics, not the other way around.

Get Early Access → See Benchmarks
01The EnginesSolver-Agnostic
PINNs
Physics-informed networks
Live
DeepONet
Operator learning
In Lab
FNO
Fourier neural operators
In Lab
Hybrid
Classical + learned
Roadmap
// The user buys the result — not the method. Every engine is validated before it ships.
02How a PINN LearnsComposite Loss
The PINN loss function
Ltotal = λdata·Ldata + λPDE·LPDE + λBC·LBC
LdataFits sparse observations — learn from a handful of sensors.
LPDE = ‖N[u] − f‖²Forces the model to satisfy the governing physics everywhere.
LBCEnforces boundary & initial conditions across ∂Ω.
03ArchitectureFully-Connected · Autodiff
Input Layer

Coords
& Time

  • x, y, z — spatial
  • t — temporal
  • continuous, mesh-free
Hidden Layers

5–8 Deep
50–200 Wide

  • sinusoidal / tanh activations
  • dense connections
  • smooth field representation
Output Layer

Solution
Field u

  • displacement
  • stress
  • temperature
AUTODIFFAutomatic differentiation computes spatial & temporal derivatives exactly — feeding the PDE residual with machine-precision gradients.
04A Training RunThermoelastic · Live
EnginePINN Standard v2
PhysicsLinear Elasticity
CollocationAdaptive
ActivationsSinusoidal
DomainWhole Assembly
StatusTraining
Training Progress
0 / 5000 epochs
36.8% completemax 5000
Residual3.2e−4
05Adaptive CollocationWhere Gradients Are High
10–100×
Faster than FEM at comparable accuracy
0 hrs
Meshing time in the loop
3.2e−4
PDE residual at current epoch
∂
Exact autodiff gradients, machine precision
// Sampling points self-concentrate near stress concentrations, boundaries and material interfaces.

Train a
Model

// Early access for research labs and R&D teams. Drop your email.

No spam. Physics only. ◆ agnostix.io