Next-Gen Neural Mass Models (NG-NMMs)

• NG-NMMs are exact low-dimensional mean-field models that derive dynamic equations for population firing rate and membrane potential from spiking neuron networks.
• They incorporate synaptic kinetics and interactions between excitatory and inhibitory populations, enabling insights into oscillatory instabilities and cross-frequency coupling.
• NG-NMMs extend to whole-brain and network simulations by integrating noise, plasticity, and structured connectivity, providing a rigorous link between microscopic and macroscopic neural dynamics.

Searching arXiv for core NG-NMM papers and related recent work. arxiv_search query: "next generation neural mass models"

Next-Generation Neural Mass Models (NG-NMMs), also called exact firing-rate models, are exact low-dimensional mean-field descriptions of large populations of class-I spiking neurons, notably quadratic integrate-and-fire (QIF) neurons and phase-equivalent $\theta$-neurons. Instead of postulating a static wave-to-pulse nonlinearity, they derive closed macroscopic equations for the population firing rate $r(t)$ and mean membrane potential $v(t)$—and, when synaptic kinetics are retained, synaptic variables such as $s(t)$—from the microscopic spiking dynamics under Lorentzian heterogeneity. In this sense, NG-NMMs preserve the coupling between rate, voltage, and synchrony, and furnish a mathematically exact bridge between microscopic QIF networks and mesoscopic neural population dynamics (Coombes et al., 2016, Clusella, 15 May 2026).

1. Exact mean-field foundations

The defining feature of NG-NMMs is that they arise from exact mean-field reduction rather than phenomenological closure. Classical heuristic neural mass models such as Wilson–Cowan, Jansen–Rit, and Lopes da Silva assume a static transfer function $r(t)=\Phi[I(t)]$ together with synaptic filters. NG-NMMs replace that assumption with dynamical equations obtained from networks of QIF neurons via the Ott–Antonsen/Watanabe–Strogatz framework and the Montbrió–Pazó–Roxin reduction. Because QIF neurons are the normal form of class-I excitability and are phase-equivalent to $\theta$-neurons, the same reduction underlies both descriptions (Coombes et al., 2016, Clusella et al., 2022).

In the standard single-population form, the exact macroscopic variables are the firing rate $r(t)$ and the mean membrane potential $v(t)$:

$$\tau \dot r = \frac{\Delta}{\pi \tau} + 2rv,\qquad \tau \dot v = \eta + v^2 - (\pi \tau r)^2 + I(t).$$

Here $\eta$ is the mean excitability, $r(t)$0 is the width of the Lorentzian distribution of neuronal currents, and $r(t)$1 is common input. This formulation does not postulate a sigmoid. In the infinitely slow synapse limit, however, the exact theory reduces to a static input–output relation,

$r(t)$2

which clarifies that the heuristic transfer function is an asymptotic approximation of the exact dynamical theory rather than its starting point (Clusella et al., 2022).

2. State variables, synaptic operators, and canonical equations

The minimal NG-NMM tracks $r(t)$3 and $r(t)$4, but most applications add explicit synaptic variables. For an inhibitory population with exponentially decaying synapses, the exact mean-field equations are

$r(t)$5

The additional variable $r(t)$6 represents the synaptic gating or field variable, and $r(t)$7 is the synaptic decay time constant. This form is especially important for interneuronal network gamma (ING), because finite synaptic kinetics can themselves generate collective oscillations (Segneri et al., 2020).

For coupled excitatory and inhibitory populations, NG-NMMs retain a separate pair $r(t)$8 and synaptic state $r(t)$9 for each population:

$v(t)$0

This two-population formulation supports both pyramidal–interneuron network gamma (PING) and ING realizations. When synapses are instantaneous, one recovers the approximation $v(t)$1. When second-order synaptic filtering is required, the synaptic operator can be written as

$v(t)$2

which is the exact counterpart of the classical second-order neural mass formalism in the slow-synapse limit (Segneri et al., 2020, Clusella et al., 2022).

3. Oscillatory mechanisms, resonance, and cross-frequency coupling

A central result of the NG-NMM literature is that collective rhythms are organized by bifurcation structure rather than imposed by transfer functions. In the exact model with second-order synapses, inhibitory QIF populations can undergo a supercritical Hopf bifurcation and produce self-sustained gamma-like oscillations, whereas the heuristic slow-synapse approximation does not reproduce this regime. Likewise, in excitatory populations the exact model admits resonant oscillatory activity whose peak frequency and amplitude increase with self-coupling gain and external excitatory input; the corresponding heuristic model yields only nodes for $v(t)$3 and therefore no resonance (Clusella et al., 2022).

Theta–gamma cross-frequency coupling is one of the best studied NG-NMM applications. In inhibitory master–slave configurations with different synaptic time scales, the exact mean field exhibits damped oscillations toward a stable focus, periodic and quasi-periodic oscillations, and chaos, and supports both phase–phase locking and phase–amplitude coupling between theta and gamma rhythms (Ceni et al., 2019). In PING and ING NG-NMMs driven near Hopf bifurcation by a half-wave shifted sinusoid,

$v(t)$4

theta-nested gamma oscillations emerge for forcing frequencies in the range $v(t)$5 Hz. In the PING set-up, a supercritical Hopf in $v(t)$6 at $v(t)$7 generates gamma limit cycles with $v(t)$8 Hz; in the ING set-up, a supercritical Hopf at $v(t)$9 yields $s(t)$0 Hz. The resulting mixed rhythms display phase–amplitude coupling, organize into 1:1 and $s(t)$1 locked states, quasi-periodic regimes, and, in ING, a chaotic window at high forcing. The locked states are more frequent in the ING set-up. The deterministic model reproduces PAC across the theta band, but the dominant gamma peak does not shift upward with $s(t)$2 unless noise amplitude or forcing amplitude is increased together with theta frequency (Segneri et al., 2020).

A complementary recent development is an energy-based interpretation of fast-synapse NG-NMMs. After the rescaling $s(t)$3, $s(t)$4, $s(t)$5 and a state-dependent time transformation $s(t)$6, the system becomes Hamiltonian in the limit $s(t)$7, with energy-like quantity

$s(t)$8

For small but nonzero heterogeneity, this quantity is no longer conserved, but it still explains why phase portraits are near-ellipsoidal and predicts spike amplitude during bursting dynamics (Andrean et al., 12 Sep 2025).

4. Network, field, and whole-brain extensions

NG-NMMs generalize naturally from single populations to structured networks. In sparse Erdős–Rényi networks of exact firing-rate nodes with row-normalized connectivity,

$s(t)$9

row normalization guarantees a homogeneous asynchronous manifold and permits a dispersion relation linking dynamical growth rates to the spectrum of the connectivity matrix. In undirected networks, only inhibitory systems produce heterogeneous stationary patterns, corresponding to a winner-takes-all mechanism. In directed networks, both excitatory and inhibitory systems can undergo Hopf-like instabilities to fast rhythmic states, and numerical simulations show high-dimensional chaos with extensive properties (Clusella, 15 May 2026).

Spatially continuous extensions produce next-generation neural field models. On a ring, the local order parameter $r(t)=\Phi[I(t)]$0 obeys a complex Riccati equation coupled through a convolution kernel, and periodic solutions satisfy an explicit self-consistency condition. That framework supports stable time-periodic solutions, provides Floquet-based stability analysis, and extends to delayed systems, two-population architectures, and networks of Winfree oscillators (Laing et al., 2023).

Large-scale connectome models inherit the same structure. In a 90-region brain network with each node implemented as an exact excitatory–inhibitory NG-NMM with exponential synapses, row-normalized anatomical connectivity preserves a homogeneous invariant manifold and allows a master-stability-style decomposition of non-uniform perturbations. The resulting dispersion relations explain the onset of traveling waves, clustered oscillations, chimera-like heterogeneous rhythms, and high-dimensional chaos. In this whole-brain setting, anatomical coupling can also generate gamma oscillations with amplitudes modulated by slower rhythms, producing sidebands and slow envelopes without adding an explicit slow oscillator to each node (Delicado et al., 3 Dec 2025).

5. Stochasticity, plasticity, and computational inference

Exact mean-field reduction does not remove finite-size effects. In NG-NMMs with instantaneous synapses, finite network size introduces shot noise, so that the mean-voltage equation becomes

$r(t)=\Phi[I(t)]$1

The corresponding shot-noise spectrum is colored rather than white,

$r(t)=\Phi[I(t)]$2

and the coupled spectrum is shaped by the exact susceptibility

$r(t)=\Phi[I(t)]$3

Even when the noise amplitude scales as $r(t)=\Phi[I(t)]$4, resonance effects can make its impact on collective dynamics crucial (Klinshov et al., 2022).

Short-term synaptic plasticity extends NG-NMMs into a synaptic theory of working memory. In the exact working-memory model, the firing rate and mean membrane potential are coupled to Tsodyks–Markram facilitation and depression variables:

$r(t)=\Phi[I(t)]$5

This formulation reproduces stimulus-locked transient oscillations followed by steady-state activity in the $r(t)=\Phi[I(t)]$6–$r(t)=\Phi[I(t)]$7 band, supports either synaptic reactivation or persistent activity, and yields an analytic approximation for maximal memory capacity,

$r(t)=\Phi[I(t)]$8

with capacity maximized for an optimal presentation-frequency range (Taher et al., 2020).

The computational ecosystem around NG-NMMs includes both high-dimensional simulation and inverse modeling. Local Linearization combined with a sparse three-dimensional Connectome Tensor enables efficient simulation of very high-dimensional neural mass networks with distributed delays, including demonstrations with 1000 ZJR columns and 3000 masses (González-Mitjans et al., 2020). Data-driven state-space reconstruction has been demonstrated with a bidirectional LSTM filter trained on simulated EEG, using a physics-linked loss that combines state reconstruction, observation reconstruction, and ODE residual consistency; simulated tests yielded correlations with $r(t)=\Phi[I(t)]$9 around $\theta$0, and GPU inference on one hour of data required approximately $\theta$1 s (Liu et al., 2023). For stochastic multi-population models, an adjusted SMC-ABC scheme jointly infers continuous biophysical parameters and binary directed connectivity from multichannel EEG using spectral densities, coherence, cross-correlation, and invariant-density summaries (Ditlevsen et al., 2023).

6. Relation to classical neural mass models, misconceptions, and limitations

A common misconception is that NG-NMMs differ from classical neural mass models only by replacing one transfer function with another. The exact comparison with second-order synapses shows a stronger claim: the heuristic NMM1 description is the infinitely slow synapse approximation of the exact NMM2. When membrane and synaptic timescales are comparable, the heuristic model fails to reproduce self-sustained oscillations of an inhibitory interneuron QIF network and also misses resonant oscillatory activity in excitatory populations. Within the broader “Rosetta Stone” classification of neural mass models, QIF-based NG-NMMs occupy the rung where the transfer function becomes dynamic rather than static, and where the parameters $\theta$2, $\theta$3, and $\theta$4 determine state-dependent gain and stimulation response (Clusella et al., 2022, Castaldo et al., 4 Dec 2025).

A second misconception is that exactness implies intrinsically oscillatory local nodes. In its simplest form, a single exact firing-rate population only displays fixed-point activity; oscillations can instead arise from synaptic kinetics, mixed excitatory–inhibitory circuitry, forcing near Hopf bifurcations, or network topology (Clusella, 15 May 2026). The exactness itself is conditional: Lorentzian heterogeneity is the mathematical device that enables closure; many studies assume fully coupled or row-normalized connectivity; noise is often modeled as additive white forcing in voltage equations; some PING analyses use instantaneous synapses; and whole-brain studies often take identical local parameters across nodes. Sparse or circuit-specific connectivity, non-Lorentzian heterogeneity, conduction delays, structured noise, and parameter sensitivity can all alter bifurcation thresholds, locking tongues, and spectral content, so generalization across brain areas requires care (Segneri et al., 2020, Delicado et al., 3 Dec 2025).

A plausible implication is that NG-NMMs are best regarded not as a single model but as a family of exact macroscopic reductions whose common advantage is rigorous micro-to-meso closure. Across this family, the recurring theme is that population firing rate, mean membrane potential, and synaptic state are treated as coupled dynamical observables rather than heuristic summaries. That shift is what allows NG-NMMs to reproduce oscillatory instabilities, cross-frequency coupling, energy-like structure, plastic working memory, topology-driven chaos, and data-assimilation workflows within one coherent formalism.