Simulating the Human Brain: A Verified Quantitative Framework

Human-scale neuron counts are no longer the bottleneck in brain emulation research. A distributed simulation with 86 billion single-compartment, leaky-integrate-and-fire neurons and 47.8 trillion synapses has already been run on 14,012 GPUs [2]. What has not been demonstrated — by that project or any other — is that a biologically faithful, individual-specific reconstruction can be measured, stored and simulated at that scale.

The open problem is information sufficiency: which structural, state, and parameter variables must be captured from a real nervous system, at what resolution, for a digital model to reproduce that system's causal responses to novel inputs and interventions, without being refit after every new experiment.

This document re-derives, checks, and extends the quantitative case for that framing. Four findings survive independent verification and are the load-bearing results of the analysis:

The table below records the outcome of independently checking every quantitative and citational claim in the prior draft against primary sources and against a from-scratch recomputation. This is the step that distinguishes a research framework from a plausible-sounding narrative: a number is not evidence until someone has tried to break it.

Full source list with live links provided in §17 (References).

The organism-ladder framing — “scale from worm, to fly, to zebrafish, to mouse, to human” — treats neuron count as the master difficulty variable. Isaak Freeman's 2026 MIT thesis develops that framing rigorously and is the direct inspiration for this line of work [1]. But neuron count is a poor proxy for emulation difficulty: a wiring diagram is extremely informative, and it still does not uniquely determine what a circuit does.

The clearest direct evidence for this comes from a connectome-constrained model of the fly visual system: with 45,669 neurons and 1,513,231 measured connections fully mapped, the modellers still had to optimise unmeasured neuronal and synaptic parameters against 26 independent functional datasets before the model predicted real responses — and the authors explicitly note that electrical synapses, neuromodulation, nonlinear chemical synapses, and glial effects were left outside the simplified model entirely [5]. Structure narrowed the solution space enormously; it did not remove the need for a dynamic-parameter inference step.

Even the best-characterised nervous system in biology makes this point starkly. The adult C. elegans hermaphrodite has a famously complete, fully mapped 302-neuron connectome. Yet the most detailed integrated brain–body–environment model built from it, BAAIWorm, models only 136 of those 302 neurons, with connection weights and polarities optimised to reproduce realistic locomotor behaviour rather than read directly off the anatomy [7]. If a 302-neuron animal with a solved connectome still requires this much inference, neuron count is not tracking the actual difficulty of emulation.

“Human-brain emulation” is defined operationally, not phenomenologically, to keep the research falsifiable: a digital model of an individual brain that reproduces and predicts its neural and behavioural responses to new inputs and interventions, within explicitly defined error bounds, without being refitted after every new experiment. Whether such a model would share the subject's subjective experience or personal identity is a separate philosophical question that this framework does not, and should not claim to, resolve.

Rather than an organism ladder, the research is organised around five measurable questions. Smaller organisms become experiments used to discover which information is sufficient, rather than rungs on a predetermined path — which gives this framework a research thesis of its own, independent of any single prior roadmap.

A cross-project benchmark compiled from primary studies, verified in §2. The purpose is not to rank organisms by neuron count; it is to compare the different kinds of information modern neuroscience can actually acquire at scale today.

Tracked fields for future comparability:volume imaged, spatial resolution, raw bytes per mm³, reconstruction completeness, detected synapses, proofreading completeness, functional coverage, temporal resolution, and whether molecular identity was captured. This avoids collapsing fundamentally different outputs into a single misleading “cost per neuron” statistic — a neuron whose soma was auto-segmented is not equivalent to one with a fully proofread arbor, synapses, and molecular identity.

Two independent nanoscale mammalian EM pipelines give two independent anchors for a naïve whole-brain linear extrapolation, using a reference volume of 1.2 × 10⁶ mm³ (≈1.2 L; see §2, item 10, for the sourcing caveat on this constant).

All three values were independently recomputed for this review and match the original draft exactly. Two unrelated present-day imaging pipelines converge on a 1.6–2.6 ZB range for naïve whole-human structural imagery under current-style nanoscale EM.

This is not a claim that 2 ZB is the answer — per §4.2, it is a cortex-calibrated figure, and compression, resolution, and segmentation choices can move it substantially. The finding that survives scrutiny is narrower and more defensible: zettabyte-scale source imagery is not an exotic assumption bolted onto this framework — it falls directly out of demonstrated mammalian connectomics data densities.

The 2026 zebrafish study is genuine and independently verified against its published abstract (§2, item 7). A remote-scanning light-sheet microscope achieved 200.8 volumes/second, covering ~85% of larval zebrafish brain volume by field of view, of which usable voltage traces were recovered from 12,935–19,039 neurons — about one-quarter of the ~78,000 estimated neurons in the imaged developmental stage [8].

These are two different bottlenecks, not one figure restated: 85% describes optical/spatial access; ~25% describes signal-quality and segmentation yield within that access. Collapsing them into a single “coverage” number would overstate what was actually recorded.

Its imaged field was ≈930 × 370 × 170 μm ≈ 0.0585 mm³, and a typical 35-second trial produced ~250 GB, giving a raw acquisition rate of:

The deeper point this number makes: a static structural map and a live functional recording have fundamentally different dimensionality. Structural imaging pays a large one-off spatial-data cost; functional capture repeatedly samples state through time. A research programme should not ask for “the storage requirement” as a single number — it should separate structural data (bytes) from functional data rate (bytes/second), and, per §4.2, should not implicitly assume either is spatially uniform across the brain.

Publishing a headline EFLOP/s figure without the constants behind it is not independently reproducible. Every parameter below is therefore listed beside the result it produces.

Anchors used: N = 86.1×10⁹ (Azevedo et al., §2 item 1); S = 47.8×10¹² (the count actually implemented in the Digital Brain simulation, §2 item 2 — a computational scale anchor, explicitly not presented as a measured human synapse count); F_n = 5 ops/ms for basic integrate-and-fire and 1,200 ops/ms for Hodgkin–Huxley (Izhikevich, §2 item 8, confirmed exact, not rounded). Event-processing constants (c_e, f_q, c_q) are stated assumptions, not measurements.

Every row was recomputed from first principles for this review and matches the original draft to the last significant figure. The arithmetic requirement changes by more than five orders of magnitude depending entirely on what a simulated neuron represents.

This reframes the central research question away from “how many FLOPs does the human brain require” — a question with no single correct answer — toward the answerable one: what is the least-complex neuron model that remains causally sufficient for the cognitive functions being reproduced?

The Digital Brain project is genuinely strong evidence that human-scale neuron counts can be distributed across large GPU systems: 86 billion neurons and 47.8 trillion synapses across 14,012 GPUs is a real, published result [2].

But verification for this review confirmed a detail the original draft left implicit: those 86 billion units were simulated as single-compartment leaky integrate-and-fire neurons — by the scenario table above, the cheapest possible model class, at roughly 0.001–0.05 EFLOP/s. This is not a criticism of that project, which was not attempting biophysical fidelity.

It is, however, decisive for this framework's argument: running 86 billion nodes and running 86 billion neurons at whatever biological fidelity ultimately proves necessary for individual-specific emulation are not the same achievement, and the gap between them, by the scenario table, can be more than five orders of magnitude in arithmetic demand alone — before memory traffic (§11) is even considered.

A single EFLOP/s headline number omits two other quantities that a systems-level treatment needs to report separately: persistent state and data movement.

These are not predictions of real cluster bandwidth — locality, sparsity, compression, caching and partitioning change physical traffic enormously — but they demonstrate why compute, memory capacity, and memory movement are three separate quantities, not one.

The Digital Brain project itself required a dedicated two-level routing design purely to move spikes between GPUs [2], independent evidence that distributing synaptic communication is a first-class systems problem, not a FLOP-counting afterthought.

The elasticity implied by the scenario table (§9.1) is necessary but not sufficient: varying one parameter at a time misses interactions such as F_n × M, where a more complex neuron model combined with more compartments multiplies rather than adds to the total demand.

The original draft reported a 100,000-draw Monte Carlo analysis and a Sobol-style global sensitivity screen without publishing code or a fixed methodology, which is not independently reproducible on its own terms — so for this review, the entire analysis was re-implemented from scratch, in Python, using SciPy/SALib, with the input assumptions stated below, an explicit random seed, and full Saltelli sampling for the sensitivity indices.

“Validation by prediction” alone is too broad to be a standard. This framework separates validation into increasing levels of rigor:

The strongest standard: a candidate emulation should predict unseen interventions without parameter refitting. This is the test that distinguishes a model that has captured causal structure from one that has simply been made flexible enough to reproduce existing data.

A validation report for any candidate emulation should therefore be a matrix, not a single pass/fail gate — reporting error separately across anatomy, spontaneous dynamics, stimulus responses, perturbations, and behaviour, since a model can legitimately succeed at some of these and fail at others.

Pulling the verified evidence together produces a picture different from a simple worm-to-human trajectory.

The research supports a paper structured around its own findings rather than as a commentary on a prior roadmap:

Suggested framing for the introduction:“This work was partly inspired by Isaak Freeman's 2026 MIT thesis on scaling brain emulation, which motivated investigation of the problem from the perspective of information sufficiency, empirical acquisition limits, computational uncertainty, and causal validation [1].” The paper should then proceed on its own terms — not as an audit of that thesis, and not organised around worm → fly → zebrafish → mouse → human as a predetermined ladder — using those organisms only where they answer a specific question in the information-sufficiency framework.

1. [1]Freeman, I. Scaling brain emulation: model organisms, connectomes, and simulation (MIT thesis, 2026).
2. [2]
   Duan, G. et al. Simulation and assimilation of the digital human brain. Nature Computational Science 4, 890–898 (2024). [Paper]
3. [3]
   Shapson-Coe, A. et al. A petavoxel fragment of human cerebral cortex reconstructed at nanoscale resolution. Science 384, eadk4858 (2024). [Paper]
4. [4]
   MICrONS Consortium. Functional connectomics spanning multiple areas of mouse visual cortex. Nature 635 (2025). [Paper]
5. [5]
   Connectome-constrained networks predict neural activity across the fly visual system. Nature 634 (2024). [Paper]
6. [6]
   Whole-brain annotation and multi-connectome cell typing of Drosophila. Nature 634 (2024). [Paper]
7. [7]
   An integrative data-driven model simulating C. elegans brain, body and environment interactions. Nature Computational Science (2024). [Paper]
8. [8]
   Wang, Y. et al. Voltage imaging of neurons distributed across entire brains of larval zebrafish. Nature Methods (14 Aug 2026). [Paper]
9. [9]
   Izhikevich, E. Which model to use for cortical spiking neurons? [PDF]
10. [10]
    UNESCO. Recommendation on the Ethics of Neurotechnology, adopted 11 November 2025, 43rd General Conference, Samarkand. [UNESCO text]
11. [11]
    Azevedo, F.A.C. et al. Equal numbers of neuronal and nonneuronal cells make the human brain an isometrically scaled-up primate brain. J. Comp. Neurol. 513, 532–541 (2009); regional breakdown from Herculano-Houzel, S. The human brain in numbers: a linearly scaled-up primate brain. Front. Hum. Neurosci. (2009). [PMC]
12. [12]
    Goriely, A. et al. Re-examination of human neuron counts and their uncertainty. Brain 148(3), 689– (2025). [Oxford Academic]