Neuron Model Reference

SC-NeuroCore currently exposes 158 lazy-loaded Python model classes across 153 Python model source modules in src/sc_neurocore/neurons/models/, plus 175 Rust PyO3 model wrappers in the optional engine. Matching model classes use the same step() / reset() / get_state() interface shape where the backend implements that surface.

Fidelity status: every model here has a faithful, tested Python reference. A stricter bar — a model whose whole four-language acceleration chain (accel/{rust,go,julia,mojo}) is real and Python-parity-proven — is tracked separately on the Model Fidelity & Polyglot Status page. Polyglot-complete today (40 models): Wang-Buzsaki, FitzHugh-Nagumo, Morris-Lecar, Connor-Stevens, Hodgkin-Huxley, AdEx, ExpIF, Lapicque, Perfect Integrator, Quadratic IF, Theta, DPI, COBA LIF, Escape Rate, Poisson, IQIF, McCulloch-Pitts, McKean, Hindmarsh-Rose, FitzHugh-Rinzel, Pernarowski, Terman-Wang, Wilson-HR, Rulkov map, GLIF, Mihalas-Niebur, Medvedev map, Cazelles map, Chialvo map, Courbage-Nekorkin map, Izhikevich 2007, Ibarz-Tanaka map, Ermentrout-Kopell, Sigmoid Rate, Threshold Linear Rate, Wilson-Cowan, Wong-Wang, Jansen-Rit, Montbrió-Pazó-Roxin (ErmentroutKopellPopulation), and Resonate-and-Fire. All other models are Python-faithful with their acceleration chain still under remediation.

Quick Start

# Python backend (default)
from sc_neurocore.neurons.models import HodgkinHuxleyNeuron
hh = HodgkinHuxleyNeuron()
spike = hh.step(current=10.0)

# Rust backend (faster, identical interface)
from sc_neurocore_engine.sc_neurocore_engine import HodgkinHuxleyNeuron
hh_rs = HodgkinHuxleyNeuron()
spike = hh_rs.step(current=10.0)

Backend Selection

Backend	Import path	Use case
Python	sc_neurocore.neurons.models	Prototyping, parameter tuning, debugging
Rust	sc_neurocore_engine.sc_neurocore_engine	Production, benchmarks, batch simulation

Backends use identical class names where parity wrappers exist (for example, HodgkinHuxleyNeuron). The Rust engine provides 175 Rust PyO3 model wrappers, 162 of which are wired into the NetworkRunner pipeline.

The package-level sc_neurocore.neurons facade remains lazy: core neuron symbols are available immediately, while model classes are resolved on first attribute access. Setting SC_NEUROCORE_NO_RUST=1 forces the pure-Python model registry even when the optional Rust engine is installed, and the fallback path caches the resolved Python class for later imports.

Rust Binding Coverage Map

tests/test_rust_python_neuron_parity.py is the live coverage map for the Python registry exposed by sc_neurocore.neurons.models.__all__. The generated capability inventory above counts 158 static classes in src/sc_neurocore/neurons/models/*.py; the registry-level map covers 160 public Python registry names because HybridFisherPosnerLIFNeuron and StochasticLIFNeuron are re-exported (from sc_neurocore.quantum_cognition.fisher_posner for population dispatch and sc_neurocore.neurons.stochastic_lif for the torch-free public root import, respectively) rather than defined as static classes under neurons/models/.

Current binding disposition:

The current registry map records 146 same-name Rust constructors, 9 Rust-prefixed or core-only constructors, and 5 Python-only registry names.

Disposition	Count	Contract
Same-name Rust constructors	146	Python registry name matches the compiled sc_neurocore_engine.sc_neurocore_engine class name.
Rust-prefixed or core-only constructors	9	A Rust binding exists, but the generic scalar parity harness uses an explicit name map.
Python-only registry names	5	No same-name PyO3 neuron constructor is claimed; each entry below records the durable boundary.

Rust-prefixed or core-only constructor map:

Python registry name	Rust constructor
AdaptiveThresholdMoENeuron	RustAdaptiveThresholdMoENeuron
AstrocyteLIFNeuron	RustAstrocyteLIFNeuron
CochlearHairCell	RustCochlearHairCell
ContinuousAttractorNeuron	RustContinuousAttractorNeuron
DendriticNMDANeuron	RustDendriticNMDANeuron
DirectionSelectiveRGC	RustDirectionSelectiveRGC
HybridLinearAttentionNeuron	RustHybridLinearAttentionNeuron
MulticompartmentMCNNeuron	RustMulticompartmentMCNNeuron
QuantumInspiredLIFNeuron	RustQuantumInspiredLIFNeuron

Python-only boundary rationale:

Python registry name	Boundary
AstrocyteNeuron	Population adapter over AstrocyteModel; the Rust engine already exposes AstrocyteModel and RustAstrocyteLIFNeuron, but this adapter threshold-converts cytosolic Ca²⁺ release into the population spike interface.
ChayKeizerMinimalNeuron	The reduced three-state pancreatic beta-cell model is a separate published reduction from the existing five-state Chay-Keizer Rust kernel and needs its own parameter-faithful implementation before a PyO3 constructor is claimed.
HybridFisherPosnerLIFNeuron	The population-compatible entry depends on the Python SpinPoolMPS quantum-metabolic state and shared-pool measurement side effects, so a same-name Rust constructor would overclaim parity until that state model is ported.
Izhikevich2007Neuron	This model already has a function-level compiled accelerator for the RK4 simulate() path (py_izhikevich2007_simulate), but no stateful PyO3 neuron constructor is claimed for the full Python object.
SRM0Neuron	The maintained Python model is an exact-flow SRM0 membrane accumulator; it is not equivalent to the older Rust SpikeResponseNeuron kernel, so the registry name stays Python-only until a faithful Rust SRM0 kernel lands.

The parity test checks this map against the Python registry, committed Rust PyO3 source declarations, and the built Rust engine when the optional engine is installed. Stochastic models without a shared RNG remain binding-covered but outside exact spike-train parity. EscapeRateNeuron and PoissonNeuron are the enrolled exceptions: their Python, Rust, Julia, Go, Mojo, schema, and RTL surfaces share one explicit seeded LFSR16 event-stream contract. IntegerQIFNeuron is the deterministic integer counterpart: every maintained runtime and Q32.0 RTL form reproduces the pinned Wu et al. source trajectory bit-for-bit. McCullochPittsNeuron preserves the original 1943 logical rule: a fixed positive active-excitatory-afferent count threshold, absolute veto by any active inhibitory afferent, and no internal cell state; signed Q32.0 uses -1 only as the hardware inhibition sentinel. ChayKeizerNeuron remains an expected parity divergence because the Python model is the five-dimensional Chay-Keizer burster while the current Rust kernel is the older reduced form.

Reference Trace Validation

The schema-driven validation harness in sc_neurocore.neurons.reference_traces loads committed JSON corpus entries, executes the UniversalNeuron runner, and compares scalar trace features with explicit per-feature tolerances. The corpus contains analytic, independently integrated, and map-iteration references tied to each entry's stated source. For example, the Ermentrout-Kopell hand-class enrolment independently advances the sourced theta flow with the maintained Euler, event, and circular-wrap conventions over 2,000 steps. A corpus entry is model-specific evidence, not a claim that every class below has external simulator parity. The Wong-Wang enrolment separately re-derives all four Euler/Ornstein-Uhlenbeck states and both pre-update rates from the 2006 Appendix and pins the paper-versus-author-code timestep discrepancy. The Jansen–Rit enrolment independently advances all six equation-(6) states, pins the published C1/C2/C3/C4 connectivity placement and Brian2 source commit, and records that the maintained 0.1 ms Euler step is implementation scope rather than a solver prescribed by the continuous paper equations. The ErmentroutKopellPopulation enrolment corrects that legacy public name to Montbrió, Pazó, and Roxin (2015), restores the printed dimensionless equation-(12) variables through R=tau*r and t'=t/tau, independently advances both states, and records that simultaneous explicit Euler is maintained implementation scope rather than a solver prescribed by the continuous paper. The ResonateAndFireNeuron enrolment independently evaluates the exact constant-input flow of Izhikevich's (2001) complex resonator, preserves x as current-like and y as voltage-like, thresholds only an upward sampled crossing of y, and checks the generalised source reset z=i*threshold rather than a radius test or origin reset.

Model Catalogue

The descriptor corpus lives under src/sc_neurocore/neurons/model_descriptors/<ClassName>.toml. Public catalogue helpers accept only public Python class identifiers such as AdExNeuron; dotted, empty, private, and path-like names fail before filesystem access. Studio uses the same load_descriptor() surface when it renders model detail pages, so the descriptor guard protects the browser-facing catalogue and maintenance tools.

tools/generate_model_descriptors.py refreshes the corpus through generate_descriptor_payload() and merge_descriptor_payloads(). The generator uses the same public-class-name boundary as the catalogue: invalid identifiers fail before registry lookup, while valid but unregistered names remain registry misses. Missing legacy v1 schemas are allowed so new models can receive honest empty curation fields, but malformed curated schemas abort the refresh instead of being silently discarded.

The v2 descriptor parser accepts both compact legacy scalar forms and expanded tables, then normalises them into ModelDescriptor, ParameterSpec, StateVariableSpec, BackendSupport, and reproducibility records. It rejects missing metadata, non-table sections, scalar tag fields, non-numeric values, malformed ranges, invalid provenance years, and invalid digest formats before a descriptor reaches the catalogue or Studio browser surface.

Tier-3 golden traces remain exact-hash contracts. When a scalar transcendental has two measured NumPy SIMD results on heterogeneous x86 hosts, the descriptor's [reproducibility] table may add golden_trace_sha256_variants; the typed parser requires every entry to be a unique, non-primary, lowercase SHA-256. The reproducibility gate accepts only that finite allowlist, reruns variant-bearing models with AVX-512 disabled, and requires the native and portable traces to stay within the same 1e-9 bound used for ULP-bounded backend parity. This records platform-level rounding honestly without turning the golden gate into a broad approximate comparison. Studio model-detail responses expose the typed variant list so clients can verify either measured platform trace without parsing the raw TOML descriptor.

Trivial IF Variants (18 models)

Python Class	Rust Class	Reference
QuadraticIFNeuron	QuadraticIFNeuron	Latham et al. 2000
ThetaNeuron	ThetaNeuron	Ermentrout & Kopell 1986
PerfectIntegratorNeuron	PerfectIntegratorNeuron	—
GatedLIFNeuron	GatedLIFNeuron	—
NonlinearLIFNeuron	NonlinearLIFNeuron	Touboul & Brette 2008
SFANeuron	SFANeuron	Benda & Herz 2003
MATNeuron	MATNeuron	Kobayashi et al. 2009
EscapeRateNeuron	EscapeRateNeuron	Gerstner 2000
KLIFNeuron	KLIFNeuron	Eshraghian et al. 2021
InhibitoryLIFNeuron	InhibitoryLIFNeuron	—
ComplementaryLIFNeuron	ComplementaryLIFNeuron	—
ParametricLIFNeuron	ParametricLIFNeuron	Fang et al. 2021
NonResettingLIFNeuron	NonResettingLIFNeuron	Brette 2004
AdaptiveThresholdIFNeuron	AdaptiveThresholdIFNeuron	Platkiewicz & Brette 2010
SigmaDeltaNeuron	SigmaDeltaNeuron	—
EnergyLIFNeuron	EnergyLIFNeuron	Sengupta et al. 2013
IntegerQIFNeuron	IntegerQIFNeuron	Wu et al. 2021, DOI 10.1109/AICAS51828.2021.9458572
ClosedFormContinuousNeuron	ClosedFormContinuousNeuron	Hasani et al. 2022

Simple Spiking (21 models)

Python Class	Rust Class	Reference
FitzHughNagumoNeuron	FitzHughNagumoNeuron	FitzHugh 1961
MorrisLecarNeuron	MorrisLecarNeuron	Morris & Lecar 1981
HindmarshRoseNeuron	HindmarshRoseNeuron	Hindmarsh & Rose 1984
ResonateAndFireNeuron	ResonateAndFireNeuron	Izhikevich 2001
BalancedResonateAndFireNeuron	BalancedResonateAndFireNeuron	Higuchi et al. 2024
FitzHughRinzelNeuron	FitzHughRinzelNeuron	Rinzel 1987
McKeanNeuron	McKeanNeuron	McKean 1970
TermanWangOscillator	TermanWangOscillator	Terman & Wang 1995
BendaHerzNeuron	BendaHerzNeuron	Benda & Herz 2003
AlphaNeuron	AlphaNeuron	—
COBALIFNeuron	COBALIFNeuron	Brette et al. 2007
GutkinErmentroutNeuron	GutkinErmentroutNeuron	Gutkin & Ermentrout 1998
WilsonHRNeuron	WilsonHRNeuron	Wilson 1999
ChayNeuron	ChayNeuron	Chay 1985
ChayKeizerNeuron	ChayKeizerNeuron	Chay & Keizer 1983
ShermanRinzelKeizerNeuron	ShermanRinzelKeizerNeuron	Sherman et al. 1988
ButeraRespiratoryNeuron	ButeraRespiratoryNeuron	Butera et al. 1999
EPropALIFNeuron	EPropALIFNeuron	Bellec et al. 2020
SuperSpikeNeuron	SuperSpikeNeuron	Zenke & Ganguli 2018
LearnableNeuronModel	LearnableNeuronModel	—
PernarowskiNeuron	PernarowskiNeuron	Pernarowski 1994

Discrete Maps (7 models)

Python Class	Rust Class	Reference
ChialvoMapNeuron	ChialvoMapNeuron	Chialvo 1995
RulkovMapNeuron	RulkovMapNeuron	Rulkov 2001
IbarzTanakaMapNeuron	IbarzTanakaMapNeuron	Ibarz, Tanaka, Sanjuan & Aihara 2007, Eqs. 2–3
MedvedevMapNeuron	MedvedevMapNeuron	Medvedev 2005 slow-calcium first-return reduction
CazellesMapNeuron	CazellesMapNeuron	Cazelles et al. 2001
CourageNekorkinMapNeuron	CourageNekorkinMapNeuron	Courbage, Nekorkin & Vdovin 2007
ErmentroutKopellMapNeuron	ErmentroutKopellMapNeuron	Ermentrout & Kopell 1986 (maintained Euler map)

Biophysical / Conductance-Based (20 models)

Python Class	Rust Class	Reference
HodgkinHuxleyNeuron	HodgkinHuxleyNeuron	Hodgkin & Huxley 1952
TraubMilesNeuron	TraubMilesNeuron	Traub & Miles 1991
WangBuzsakiNeuron	WangBuzsakiNeuron	Wang & Buzsáki 1996
ConnorStevensNeuron	ConnorStevensNeuron	Connor et al. 1977
DestexheThalamicNeuron	DestexheThalamicNeuron	Destexhe et al. 1993
HuberBraunNeuron	HuberBraunNeuron	Braun et al. 1998
GolombFSNeuron	GolombFSNeuron	Golomb et al. 2007
PospischilNeuron	PospischilNeuron	Pospischil et al. 2008
MainenSejnowskiNeuron	MainenSejnowskiNeuron	Mainen & Sejnowski 1996
DeSchutterPurkinjeNeuron	DeSchutterPurkinjeNeuron	De Schutter & Bower 1994
PlantR15Neuron	PlantR15Neuron	Plant & Kim 1976
PrescottNeuron	PrescottNeuron	Prescott et al. 2008
MihalasNieburNeuron	MihalasNieburNeuron	Mihalas & Niebur 2009
GLIFNeuron	GLIFNeuron	Allen Institute GLIF5
GIFPopulationNeuron	GIFPopulationNeuron	Mensi et al. 2012
AvRonCardiacNeuron	AvRonCardiacNeuron	Av-Ron et al. 1991
DurstewitzDopamineNeuron	DurstewitzDopamineNeuron	Durstewitz et al. 2000
HillTononiNeuron	HillTononiNeuron	Hill & Tononi 2005
BertramPhantomBurster	BertramPhantomBurster	Bertram et al. 2000
YamadaNeuron	YamadaNeuron	Yamada et al. 1989

Multi-Compartment (7 models)

Python Class	Rust Class	Reference
PinskyRinzelNeuron	PinskyRinzelNeuron	Pinsky & Rinzel 1994
HayL5PyramidalNeuron	HayL5PyramidalNeuron	Hay et al. 2011
MarderSTGNeuron	MarderSTGNeuron	Marder & Calabrese 1996
RallCableNeuron	RallCableNeuron	Rall 1964
BoothRinzelNeuron	BoothRinzelNeuron	Booth et al. 1997
DendrifyNeuron	DendrifyNeuron	Beniaguev et al. 2022
TwoCompartmentLIFNeuron	TwoCompartmentLIFNeuron	—

Stochastic / Population / Neural Mass (13 models)

Python Class	Rust Class	Reference
PoissonNeuron	PoissonNeuron	Gerstner et al. 2014, Sections 7.2 and 7.7
InhomogeneousPoissonNeuron	InhomogeneousPoissonNeuron	—
GammaRenewalNeuron	GammaRenewalNeuron	—
StochasticIFNeuron	StochasticIFNeuron	—
GalvesLocherbachNeuron	GalvesLocherbachNeuron	Galves & Löcherbach 2013
SpikeResponseNeuron	SpikeResponseNeuron	Gerstner 1995 (SRM0)
GLMNeuron	GLMNeuron	Pillow et al. 2008
WilsonCowanUnit	WilsonCowanUnit	Wilson & Cowan 1972
JansenRitUnit	JansenRitUnit	Jansen & Rit 1995
WongWangUnit	WongWangUnit	Wong & Wang 2006
ErmentroutKopellPopulation	ErmentroutKopellPopulation	Montbrió et al. 2015
WendlingNeuron	WendlingNeuron	Wendling et al. 2002
LarterBreakspearNeuron	LarterBreakspearNeuron	Breakspear et al. 2003

Hardware Chip Emulators (9 models)

Python Class	Rust Class	Reference
LoihiCUBANeuron	LoihiCUBANeuron	Davies et al. 2018 (Intel Loihi)
Loihi2Neuron	Loihi2Neuron	Intel Loihi 2
TrueNorthNeuron	TrueNorthNeuron	Merolla et al. 2014 (IBM)
BrainScaleSAdExNeuron	BrainScaleSAdExNeuron	Schemmel et al. 2010
SpiNNakerLIFNeuron	SpiNNakerLIFNeuron	Furber et al. 2014
SpiNNaker2Neuron	SpiNNaker2Neuron	TU Dresden 2024
DPINeuron	DPINeuron	Indiveri, Stefanini & Chicca 2010
AkidaNeuron	AkidaNeuron	BrainChip
NeuroGridNeuron	NeuroGridNeuron	Boahen 2014

Rate / Plasticity / Other (12 models)

Python Class	Rust Class	Reference
McCullochPittsNeuron	McCullochPittsNeuron	McCulloch & Pitts 1943
SigmoidRateNeuron	SigmoidRateNeuron	Wilson & Cowan 1972
ThresholdLinearRateNeuron	ThresholdLinearRateNeuron	—
AstrocyteModel	AstrocyteModel	Li & Rinzel 1994
TsodyksMarkramNeuron	TsodyksMarkramNeuron	Tsodyks & Markram 1997
LiquidTimeConstantNeuron	LiquidTimeConstantNeuron	Hasani et al. 2021
CompteWMNeuron	CompteWMNeuron	Compte et al. 2000
SiegertTransferFunction	SiegertTransferFunction	Siegert 1951
FractionalLIFNeuron	FractionalLIFNeuron	Teka et al. 2014
ParallelSpikingNeuron	ParallelSpikingNeuron	Fang et al. 2023
AmariNeuralField	AmariNeuralField	Amari 1977
LeakyCompeteFireNeuron	LeakyCompeteFireNeuron	—

Core IF (Rust neuron.rs) (3 models)

Python Class	Rust Class	Reference
AdExNeuron	AdExNeuron	Brette & Gerstner 2005
ExpIFNeuron	ExpIFNeuron	Fourcaud-Trocmé et al. 2003
LapicqueNeuron	LapicqueNeuron	Lapicque 1907

Common Interface

All models share:

model.step(current)   # → int (spike: 1/0) or float (firing rate)
model.reset()         # → None (restore initial conditions)
model.get_state()     # → dict of current state variables

Multi-input models accept additional arguments:

• AlphaNeuron.step(exc_current, inh_current)
• COBALIFNeuron.step(current, delta_ge, delta_gi)
• PinskyRinzelNeuron.step(current_soma, current_dend)
• TsodyksMarkramNeuron.step(current, presynaptic_spike)
• CompteWMNeuron.step(current, spike_in)

Neural mass models return float (firing rate or EEG potential):

• WilsonCowanUnit, JansenRitUnit, WendlingNeuron
• ErmentroutKopellPopulation, LarterBreakspearNeuron
• SigmoidRateNeuron, SiegertTransferFunction

WongWangUnit.step(stim1, stim2) returns the two continuous pre-update rates as (r1, r2). Its deterministic batch additionally accepts interleaved external Gaussian samples and returns six complete traces plus four final states.

JansenRitUnit.step(p_ext) returns the post-update continuous y1-y2 EEG proxy. Its atomic batch accepts one external-drive sample per Euler step and returns all six state traces, the EEG trace, and six final-state receipts.

ErmentroutKopellPopulation.step(ext_input) returns the post-update continuous population firing rate from the Montbrió–Pazó–Roxin exact QIF-network mean field. Its atomic batch returns complete r and v traces plus both final- state receipts; the compatibility name does not denote the 1986 single-cell Ermentrout–Kopell theta equation.

AI-Optimized (9 models)

Novel neuron models designed for AI workloads, not biological simulation. Located in neurons/models/ai_optimized.py and neurons/models/arcane_neuron.py.

Python Class	Rust Class	Key Feature
ArcaneNeuron	ArcaneNeuron	5-compartment self-referential cognition: fast (5ms), working memory (200ms), deep context (10s), attention gate, forward self-model. Identity accumulates in the deep compartment. Confidence modulates threshold and meta-learning rate. Sotek & Arcane Sapience 2026.
MultiTimescaleNeuron	—	Three-compartment (fast/medium/slow) with context-dependent threshold modulation
AttentionGatedNeuron	—	Learned sigmoid gate (key/query weights) selectively filters input
PredictiveCodingNeuron	—	Fires only on prediction errors (novel stimuli), silent on expected input
SelfReferentialNeuron	—	Introspects own spike history to auto-regulate firing dynamics
CompositionalBindingNeuron	—	Phase-coding for variable binding; in-phase = bound concepts
DifferentiableSurrogateNeuron	—	Trainable surrogate gradient parameters (alpha, beta, theta)
ContinuousAttractorNeuron	—	Ring attractor with Mexican-hat connectivity for continuous working memory
MetaPlasticNeuron	—	Self-regulating meta-learning rate based on error trace