---

name: dynamics-modeling description: "Construct and route NeuroMANCER neural blocks, continuous and stochastic dynamics, physics-informed models, integrators, interpolation, and system-identification components without mixing in training or control-loop concerns." disable-model-invocation: true metadata: disco-role: operating license: NOASSERTION

Dynamics modeling

Use this skill when the task is to construct, inspect, or compose a model of a dynamical system in NeuroMANCER. Keep the runtime model self-contained: use package imports and the contracts in the bundled references, not repository-relative imports or example files.

Route the request

• Use references/api-reference.md for stable imports, signatures, tensor shapes, and optional backends.
• Use references/workflows.md for build recipes: neural state-space/ODE models, named or SINDy systems, networked physics, DAE operator splitting, PINN residuals, function encoders, and interpolation/SDE paths.
• Use references/troubleshooting.md when dimensions, autograd, solver options, device/dtype, or optional dependencies are involved.
• Run scripts/dynamics_smoke.py --help; run scripts/dynamics_smoke.py --run in an environment where the package's ODE dependencies are installed.

Core routing rules

1. Choose the tensor contract before choosing the class. Ordinary blocks consume rank-2 [batch, features] tensors. Block.forward(*inputs) concatenates multiple rank-2 inputs along the last axis. Sequence wrappers have their own [batch, time, features] contract; see the API reference.
2. Separate state width from input width. An autonomous RHS maps [B, nx] -> [B, nx]. A non-autonomous RHS is constructed with insize=nx+nu, outsize=nx, but is called as rhs(x, u) with [B, nx] and [B, nu] tensors.
3. Use Euler, RK2, or RK4 for transparent one-step CPU integration; use DiffEqIntegrator for torchdiffeq methods and adjoint differentiation. Use the multistep and second-order families only with their specialized state layouts.
4. Treat SDEs, PINNs, and DAE models as explicit advanced paths. SDE integration requires TorchSDE. PINN coordinates must retain a differentiable autograd graph. DAE examples use algebraic-update then differential-update composition rather than a hidden generic solver.
5. Do not put losses, datasets, trainers, or control rollouts here. Route loss/data/training setup to symbolic-problems or data-training, and closed-loop/system simulation to control-simulation. Route structured-map registry selection or native extensions to structured-operators.
6. Do not promise GPU or native-extension coverage. The CPU construction and shape paths are the baseline; TorchSDE/TorchDiffEq, CUDA, and native structured operators are separately verified capabilities.

Minimal construction shape

import torch
from neuromancer.modules.blocks import MLP
from neuromancer.dynamics.integrators import RK4

nx, nu = 2, 1
rhs = MLP(nx + nu, nx, linear_map=torch.nn.Linear,
          nonlin=torch.nn.Tanh, hsizes=[16, 16])
stepper = RK4(rhs, h=0.05)
x = torch.zeros(8, nx)
u = torch.zeros(8, nu)
x_next = stepper(x, u)       # [8, 2]

For complete model variants, shape checks, and the model-versus-training boundary, follow the linked references rather than expanding this router.