Integrated Guide · Chapter 2

Scientific fundamentals — what must be copied, and at what resolution.

Whole brain emulation can be summarised in one sentence — scan a brain, run it on a computer — but that sentence decomposes into six independent engineering problems. This chapter covers the numbers that describe the object, the stages of the pipeline, the field's largest open question (deciding the required resolution), the available scanning technologies, the compute and data requirements, and the serious scientific objections.

Level: intermediate, no prerequisites28 references

In this chapter

2.1 The object: brains in numbers
2.2 The six-stage WBE pipeline
2.3 The resolution problem
2.4 Scanning technologies compared
2.5 Compute and data
2.6 Objections, and current responses
2.7 Chapter summary
References

2.1 The Object

The object: brains in numbers.

Definition

Whole brain emulation (WBE): measuring the relevant structure and state of a particular brain at sufficient resolution and reproducing its causal information processing on another computational substrate. Unlike abstract brain-inspired AI, the defining feature is reproducing one specific individual¹. Simulation (modelling brains in general) and emulation (functionally copying a particular brain) are distinct.

Organism	Neurons	Synapses (approx.)	Brain volume (approx.)	Notes
Nematode C. elegans	302	≈7×10³	— (1 mm body)	The only complete whole-animal wiring diagram²
Adult fly (brain)	139,255	≈5.5×10⁷	≈0.08 mm³	Whole-brain connectome completed 2024³
Larval zebrafish	≈1×10⁵	≈10⁸	≈0.1 mm³	Largest brain whose activity can be recorded whole
Mouse	≈7×10⁷	≈10¹¹	≈500 mm³	Structural map exists for 1 mm³ only⁴
Human	8.6×10¹⁰	≈1.5×10¹⁴	≈1.2×10⁶ mm³	Neuron count from Herculano-Houzel et al.⁵; runs on about 20 watts

What matters is not the absolute numbers but the distance in orders of magnitude. Complete structural maps reach the fly (10⁵). The mouse is about three orders further, the human three beyond that. Since data volume scales with tissue volume, a whole human brain means roughly 15 million times the imaging and processing of a whole fly brain. Working the other way, the operating principles of neurons — membrane potentials, synaptic transmission, neuromodulation — are well conserved across species, so methods established in small brains do scale in principle. "The distance is large but the road is continuous" is the basic intuition of this field.

2.2 The Pipeline

The six-stage WBE pipeline.

Since the 2008 roadmap¹, WBE has been understood as the following dependency pipeline. Each stage consumes the output of the previous one, and the whole collapses if any single one is missing.

Stage	What it does	Representative technology	Principal difficulty
1. Preserve / fix	Fix the brain's fine structure without degradation and store it	Perfusion fixation; aldehyde-stabilized cryopreservation (ASC)⁶	Racing post-mortem decay; achieving uniform perfusion across a whole brain
2. Image (scan)	Image the full volume at a resolution where synapses are visible (single to tens of nm)	Electron microscopy in its various forms; expansion microscopy with optics⁷	Throughput. At current EM rates a human brain takes centuries without massive parallelism
3. Reconstruct	Extract every neuron's shape and every synapse from the images to produce a wiring diagram	AI segmentation with human proofreading³	Proofreading cost — tens of person-years for the fly. Errors propagate
4. Model	Assign synaptic weights, transmitters and dynamical parameters to the static wiring diagram	Inference from structure, molecular annotation, digital-twin learning⁸	The greatest scientific uncertainty. How much dynamical information static structure contains is unknown
5. Run	Execute the model and couple it to a body and environment	Spiking simulators; physics-simulated bodies⁹	Human-scale real-time execution; fidelity of body and environment
6. Validate	Show that the emulation is "the same as" the original brain	Behavioural matching, perturbation response, activity-prediction benchmarks¹⁰	No agreed pass criteria. Validating personal identity is a problem in principle (Chapter 5)

How to read the news

When you see a headline about a brain-mapping breakthrough, the first useful question is which stage of this table it belongs to. Stages 2 and 3 (mapping) are advancing quickly; stages 4 (modelling) and 6 (validation) lag by orders of magnitude. That asymmetry is the whole explanation for the gap between excited coverage and cautious expert timelines.

2.3 Resolution

The resolution problem: what is enough to copy?

WBE's biggest open problem is not an instrument but a specification. There is still no experimental answer to the question of which physical quantities must be copied in order to reproduce a mind. The roadmap organised this as a hierarchy of levels¹.

Emulation level	What is copied	Compute required (approx.)	Current standing
Spiking network	Every neuron's firing and every synaptic weight	≈10¹⁸ FLOPS	Working lower bound The fly reproduced behaviourally relevant responses at this level⁹
Electrophysiological detail	Membrane potentials per dendritic compartment (Hodgkin–Huxley type)¹¹	≈10²² FLOPS	Plausible safe assumption Demonstrated for cortical microcircuits¹²
Metabolic / molecular dynamics	Receptor distributions, neuromodulators, intracellular signalling	10²⁵–10²⁷ FLOPS	Partially necessary If required in full, feasibility recedes substantially
Stochastic molecular level	The behaviour of individual molecules	≈10⁴³ FLOPS	Effectively impossible There is currently no evidence this level is needed

"The connectome is not enough" — what is missing

A wiring diagram is necessary but not sufficient. The following are either invisible in static EM images or of uncertain recoverability from them.

Synaptic weights. EM shows only morphological proxies such as contact area and vesicle counts. The mapping from morphology to functional strength is partial⁴.
Neuromodulators. Dopamine, serotonin and others reconfigure a circuit with fixed wiring into effectively different circuits depending on context — the classic lesson of crustacean stomatogastric ganglion work¹³¹⁴.
Electrical synapses and volume transmission. Gap junctions are easily missed in EM, and diffusion-based extrasynaptic signalling lies outside the wiring diagram entirely.
Glia. Roughly half of brain cells, involved in synaptic modulation and homeostasis. The precision at which they must enter an emulation is unsettled.
Plasticity and learning rules. Copying the brain at an instant is not enough; without correctly implementing the rules by which it changes, a personality will diverge over time.
Degeneracy. Different parameter sets can produce identical observed activity, so more measurement does not necessarily converge on a unique solution¹³. Narrowing candidates requires perturbation experiments and predictive validation.

The nematode paradox — the field's most important lesson

The nematode connectome has existed since 1986, yet forty years later there is no validated nematode emulation. The reasons are clear: (1) the wiring diagram carries no weights, (2) nematode neurons are largely non-spiking and electrophysiological data are scarce, and (3) the circuit depends heavily on neuromodulation²¹⁴. Meanwhile the fly, with 460 times more neurons, yielded a validated reproduction from a simple leaky integrate-and-fire model, thanks to spiking neurons and abundant behavioural data⁹. The lesson: difficulty is set not by size but by the distance between the map and the dynamics.

2.4 Scanning

Scanning technologies compared.

Technology	Resolution	Track record	Strengths / limits
Serial-section EM (ssTEM/SEM)	≈4 nm	FlyWire (whole fly brain)³	Synapses reliably visible / sectioning is destructive and slow
Multibeam SEM	≈4 nm	H01 (1 mm³ human cortex = 1.4 PB)¹⁵	Fastest available EM / still orders of magnitude short of a human brain
FIB-SEM	≈8 nm isotropic	Hemibrain¹⁶	Best 3D quality / small accessible volumes
Expansion microscopy + light sheet	effective ≈10–60 nm	Demonstrated in fly brain and cortex⁷	Fast, allows molecular labelling / weak at tracing connectivity alone
Protein barcoding (PRISM)	optical + barcodes	10⁷ μm³ of mouse hippocampus (E11 Bio)¹⁷	Designed to cut proofreading cost by orders of magnitude / large-scale demonstration still ahead
X-ray nanotomography (XNH)	≈100 nm	Mouse and fly nerve bundles¹⁸	Fast and section-free / does not reach synaptic resolution alone
DNA barcoding (MAPseq etc.)	cellular	Mouse projection maps¹⁹	Very fast and cheap / cannot resolve synapse-level wiring
MRI / diffusion MRI	≈0.5–1 mm	Whole human brain (Human Connectome Project)	Non-destructive, works in vivo / about five orders of magnitude too coarse

The non-destructive scanning barrier

Within known physics, there is no physical means of reading a living human brain at synaptic resolution. Light scatters within about a millimetre, MRI is five orders of magnitude short, and no method exists in principle for selectively reading nanometre structures at depth with electromagnetic radiation. Every realistic WBE plan therefore assumes destructive scanning of a preserved brain — which is why preservation technology and ethics become central rather than peripheral (Chapters 3 and 5). The exceptional alternative is to connect brain and machine through a long-lived interface and migrate function gradually (as in Watanabe's hemisphere-connection proposal²⁰), but the required interface bandwidth has not been achieved either.

2.5 Compute & Data

Compute and data.

Raw scan data. One cubic millimetre of human cortex is 1.4 petabytes of EM imagery¹⁵. Scaled naively, a whole human brain (about 1.2 million mm³) is 1 to 2 zettabytes — a few percent of annual global data creation in the mid-2020s. Storing it matters less than building a pipeline that compresses and reconstructs while imaging.
Runtime compute. The roadmap estimated ≈10¹⁸ FLOPS at spiking level¹. Supercomputers passed exascale (10¹⁸) in 2022²¹, so in raw arithmetic we already touch the lower-bound hypothesis for one human. An independent estimate puts the functional requirement lower still, at 10¹³–10¹⁷ FLOP/s²².
The real bottleneck. Spiking network simulation is limited by memory capacity, bandwidth and communication rather than arithmetic. Holding the state of 10¹⁴ synapses (tens to hundreds of terabytes) and updating and routing it every millisecond is harder than the FLOPS figure suggests.
Neuromorphic computing. Hardware specialised for spike communication — SpiNNaker²³, Intel's Hala Point at 1.15 billion neurons²⁴, DeepSouth at 228 trillion synaptic operations per second²⁵ — runs alongside as a candidate for low-power whole-brain execution.
AI compression as a new route. The 2025 digital-twin work showed a path that does not solve detailed biophysics explicitly, instead learning response functions from large volumes of activity data⁸. Relaxing the fidelity criterion from "copy the mechanism" to "copy the input–output relation" could cut compute by orders of magnitude — though whether the result deserves to be called that person is the subject of Chapter 5.

2.6 Objections

Objections, and current responses.

A textbook should present the case against without softening it. Each objection is given with its current status.

Objection	Representative proponent	Current response and status
"Centuries away on scale alone"	K. Miller (2015)²⁶	The shortfall in imaging and proofreading is real. But AI automation has outrun Miller's assumptions — the whole fly brain arrived earlier than his 2015 framing implied. Partly valid
"The brain is not computable"	M. Nicolelis²⁷	The claim that the brain's physical processes cannot be simulated computationally has little empirical support and conflicts with the success of quantitative neuroscience since Hodgkin–Huxley. Minority view
"Quantum effects are essential"	Penrose & Hameroff	The standard rebuttal is that quantum coherence in the brain decoheres almost instantly under thermal noise (Tegmark)²⁸. Experimental support is limited and it is not a mainstream hypothesis, though not conclusively settled either. Minority view
"The connectome is not enough"	Marder, Bargmann and others¹³¹⁴	Correct. The WBE side accepts this and has evolved toward integrated structural, molecular and activity measurement with perturbation-based validation. Less an objection than a design requirement. Accepted requirement
"Body and environment are indispensable"	The embodied-cognition tradition	Broadly agreed. The roadmap already included body and environment models, and the 2026 fly demonstration was embodied from the outset⁹. Absorbed into design
"Reproducing function does not produce consciousness"	Searle, Seth and others	Not an empirical question but a philosophical one, and unsettled: theories make opposite predictions (IIT says no, functionalism says yes). Treated in Chapter 5. Unsettled

2.7 Summary

Chapter summary.

Key points

WBE is a six-stage pipeline: preserve → image → reconstruct → model → run → validate. Mapping (2–3) is fast; modelling (4) and validation (6) are slow.
The largest open problem is not an instrument but a specification: the required resolution is undetermined, with spiking-to-electrophysiological levels as the working hypothesis. If molecular detail turns out to be necessary throughout, feasibility recedes sharply.
Compute has reached the order of magnitude of the lower-bound hypothesis. The real walls are data volume (≈1–2 ZB for a human), memory bandwidth, and automated proofreading.
Non-destructive scanning of a living brain has no visible physical path, so the realistic route is destructive scanning of preserved brains — which makes preservation and ethics central scientific questions.
"The connectome is not enough" is not an objection but an established design requirement, and the nematode paradox is its proof.

References