All medical POCs
POC 03CONFIRMED

Neural Tissue Network Simulation

A 100-neuron FitzHugh-Nagumo network on an Erdős-Rényi graph — 200 coupled state variables with synaptic noise — where standard solvers fail completely.

30/30
SolvSRK survived (200-dim)
0/30
BDF seeds survived
380ms
SolvSRK median wall time
200
State variables (dim)

The scenario

Set the picture

Computational neuroscience models neural tissue as networks of coupled neurons. The FitzHugh-Nagumo model captures the essential excitation-recovery dynamics of each neuron with two state variables (membrane potential and recovery variable). A 100-neuron network on an Erdős-Rényi random graph has 200 coupled ODEs.

In experimental neuroscience, synaptic noise, channel noise, and measurement artifacts are unavoidable. Neural network simulations must integrate through this noise to produce meaningful predictions of network synchronization, wave propagation, and emergent dynamics. This is the highest-dimension problem in the medtech validation suite.

Cost today

Standard implicit solvers fail catastrophically on this system under noise. BDF crashed in under 300ms — the noisy 200-dimensional Jacobian estimation is computationally intractable. Radau timed out at 10 seconds per seed, consuming 357K–385K function evaluations without converging.

Researchers using standard solvers on noisy neural network models must either filter noise out before integration (losing biologically relevant stochastic dynamics) or reduce network size until the solver can handle it (losing emergent network-scale phenomena).

What changes with SolvSRK

SolvSRK's noise-conditioning architecture scales to 200 state variables without the failure modes that cause standard solver crashes under noise. On the 100-neuron FHN network with Gaussian noise, SolvSRK achieved 100% survival across all 30 seeds. Median wall time: 380ms. Median function evaluations: 9,233.

This means researchers can run full-scale neural tissue simulations with biologically realistic noise — synaptic noise, channel noise, neuromodulatory fluctuations — without reducing network size or filtering out stochastic dynamics. The noise is part of the biology, not an obstacle.

Measurable outcome

What we claim — and how it survives review

Each line below maps to a captured number in the demo section. Every number is reproducible from the benchmark suite.

  • SolvSRK: 30/30 seeds survived (100% survival rate).
  • BDF: 0/30 survived — crash in <300ms, 3K–12K nfev.
  • Radau: 0/30 survived — timeout at 10s, 357K–385K nfev.
  • SolvSRK median wall time: 380ms for 200 coupled state variables.
  • This is the highest-dimension problem tested: 200 state variables, stiffness S1.
  • Sub-finding SF-MEDTECH-3: 200-dim survival with no degradation vs lower-dim problems.

The demo

What was tested. How. What the simulation printed.

Benchmark: 100-neuron FitzHugh-Nagumo network on Erdős-Rényi random graph. Stiffness class S1, dimension 200, t_span [0, 200.0]. Gaussian noise injected at sigma=0.001 on every RHS evaluation.

Three solver arms: SolvSRK, scipy BDF, scipy Radau. 30 seeds per solver, 90 total runs. Reference solution computation failed (SCD = None) — precision comparison is not available for this system. Survival verdict is based on integration completion, not accuracy comparison.

The 200-dimensional Jacobian makes implicit methods (BDF, Radau) computationally prohibitive under noise. SolvSRK avoids the failure modes that cause standard solver crashes, achieving consistent 380ms integrations regardless of noise seed.

Captured benchmark output

The numbers the simulation actually printed.

Solver comparison on 100-neuron FHN network (S1, dim=200, sigma=0.001)
SolverSurvivedSurvival %Median nfevMedian wall (s)Failure mode
SolvSRK30/30100%9,2330.380
BDF0/300%<0.3Crash
Radau0/300%10.0 (cap)Timeout

Claim ID: MEDTECH-FHN100. Topic: solvsrk-reval-medtech. 30 seeds, Gaussian noise sigma=0.001. Reference solution failed — SCD not available.

Evidence pointers

Where the claims live in the evidence register

These are the validation sources a reviewer should trace to verify every number on this page.

  • Claim MEDTECH-FHN100 — Grade A CONFIRMED. Topic: solvsrk-reval-medtech.
  • Sub-finding SF-MEDTECH-3: 200-dim FHN network survives with no degradation.
  • Problem: fhn_network_100 (Erdős-Rényi 100-neuron graph). Stiffness: S1. Dim: 200.
  • 90 runs total (30 SolvSRK + 30 BDF + 30 Radau). Noise: Gaussian sigma=0.001.
  • Reference solution FAILED — no SCD available. Survival-only verdict.
  • Triage date: 2026-05-08. Phase verdict: CLOSED.

Want to see these numbers on your model?

Run the benchmark on your actual physiological system.

Two weeks, fully credited. Every claim above traces back to a simulation you can verify.

Talk to us