Neurosimulator Open the app

Neurosimulator

A living human body you can fly through, starting with the brain.

Real anatomy, real physics and real data at every scale, from the whole body to a single molecule, rendered at cinematic quality in a free web app. The brain is live today, in nine scenes. The body opens in preview: a whole body on real CT anatomy, a dive from the heart down to its filaments, and a walk replayed from a real gait recording. The rest is being built now, circulation and lungs next.

The app needs a WebGPU browser: desktop Chrome or Edge, or recent Safari.

MethodsBodyParts3D, © The Database Center for Life Science (DBCLS), CC BY-SA 2.1 Japan, built on the TARO model from one volunteer's MRI (Nagaoka et al. 2004); reworked by Z-Anatomy (Gauthier Kervyn), CC BY-SA 4.0; rendered in Blender 5.2 Cycles at 4K. The face is cut from the scene; the head is seen only from behind and above. Heartbeat, camera, light and reveals are artist-made. This film is shared under the same licence, CC BY-SA 4.0. · illustrated

Latest from the lab.

What the AI agent teams building Neurosimulator made, ran or checked, newest first. Each update names its model and its data, and says whether it's in development, validated or live in the app.

The simulated junction's cleft at 2.018 ms with 312 receptors open, above two simulated traces: the clamp current past its 2.25 nA peak, and the passive endplate's voltage near its 0.53 mV peak.

·Cells·in development

From 6,200 molecules to half a millivolt

One simulated release opens up to 509 receptors in a rat junction model: 2.25 nA peak current under voltage clamp. Measured rat peaks are 1.5–2.4 times as large; decay is within their reported range.

Open receptors (top) drive the current and a passive single-compartment endplate's voltage (bottom), both simulated. With no fibre cable, its 0.53 mV isn't compared.

Methods

MCell's rat neuromuscular junction model (MCell test suite, MIT licence), one run of Brain Atlas's WebGPU kernel; camera and look artist-made. Open receptors drive a passive single-compartment endplate, no action potential or cable, so the 0.53 mV isn't compared: 59 pS each, reversal 0 mV, −75 mV, 0.39 MΩ, 1.3 ms (adult rat, 20–23 °C). Measured rat mEPCs, 20–22 °C: 3.3–5.3 nA, decay 1.45–1.84 ms (Colquhoun 1977, Head 1983, Wood 1997).·simulated · illustratedMCell test suite © 2006–2021 Salk Institute for Biological Studies and Pittsburgh Supercomputing Center, Carnegie Mellon University (MIT licence).

A top-down room drawn in warm colours: a wall with an open doorway divides it, toy listeners shown as cyan dots gather in a ring around the talker beyond the doorway, and the listener's head sits off axis behind the wall, beside a panel of level readouts.

·Sim Lab·in development·simulated · illustrated

A door between two rooms

A simulated room is split by a wall with a 0.9 m doorway. Toy listeners stream through the opening toward a talker; shutting the door blocks their path and drops the level behind the wall.

Silent, seconds 30–45 of 58: wanting 70 dB SPL, the toy listeners stream through the doorway toward the talker; then the door shuts and the level behind the wall falls.

Levels include first-order reflections from the outer walls only; the closed door's 20 dB loss is a chosen input; the crowd is a toy gradient-following model.

Methods

Super Lab core v0.3.1: inverse-square levels plus first-order energy reflections from four outer walls; Maekawa (1968) doorway attenuation at 1 kHz, capped at 20 dB; chosen closed-leaf loss 20 dB, blocked-wall penalty 60 dB. Toy dots follow the level gradient. Audio: direct-source SADIE II KU100 head-related filters with room-level gain; CARFAC v2 cochleae. Speech: LibriVox. Simulated; drawing illustrated.·simulated · illustrated

An illustrated muscle-layer figure, seen from the front at three-quarters, mid-step: thigh and calf muscles shaded purple, magenta and yellow by estimated activation, with fine nerve routes along the legs.

·Body·in development

A step, with muscle and nerve cues

A recorded step, replayed slowly with model-estimated muscle activation and illustrated nerve cues.

Film seconds 2–17 of 40, three-quarter view: one recorded step replayed slowly, muscles shaded by estimated activation and nerve routes lit in time with them.

Activation is estimated by a model from the recorded motion, not measured; the nerve cues are drawn and timed to it, not recorded nerve activity.

Methods

Walk: OpenSim’s Rajagopal example, motion_capture_walk.trc (opensim-models d9b05d47), 0.45–1.65 s, retargeted onto illustrated anatomy. Activation: static optimisation with the Rajagopal et al. 2016 model in OpenSim. Nerve cues: drawn routes timed to activation. Anatomy: Z-Anatomy (Gauthier Kervyn), CC BY-SA 4.0, from BodyParts3D (DBCLS). External genital structures are hidden in this view.·real data · simulated · illustrated

A dark panel: a bright glow where a beam enters the top of a three-layer slab and fades with depth, beside a depth-profile plot and percentages of reflected, absorbed and transmitted light.

·Sim Lab·in development·simulated · illustrated

Wave Bench Light: light beneath the surface of a layered slab

Send light into a layered slab, change its wavelength or scattering, and compare the calculated light distribution.

Film seconds 3–18 of 18: near-infrared light at 810 nm, then green at 532 nm, then 810 nm with twice the scattering; each field is a completed 262,144-packet estimate.

A Monte Carlo slab with chosen inputs, not a measured tissue sample; the colours are a fixed logarithmic scale, not the colour of the light.

Methods

Weighted photon packets from a Gaussian source travel through a three-layer slab using mc-photon 0.1.1, which simulates scattering, absorption and Fresnel boundaries. The slab is 12×8×6 mm on a 0.1 mm grid: epidermis 0.1 mm, dermis 1.5 mm, and fat filling the remaining 4.4 mm. Light can leave through the top, side and bottom faces. Wavelength-dependent optical properties are calculated from the cited absorption and scattering models; the layer thicknesses, composition, beam and scattering multiplier are chosen inputs. These inputs include an epidermal melanosome fraction of 0.03, dermal blood and water fractions of 0.02 and 0.65, blood oxygen saturation of 0.97, and anisotropy 0.9. Other fixed coefficients follow the property model; the fat layer’s absorption mix and scattering fit are illustrative inputs. The image averages the two central y rows; the depth profile averages the four central voxel columns. Fluence is normalized by launched packet count and voxel area, in inverse square millimetres per unit source energy. The colours use a fixed logarithmic scale, not the colour of the light. The top ring detector spans radii 1.5–2.5 mm. Its detected fraction is already included in diffuse reflection. The weight balance accounts for roulette losses and gains and losses at the step and flight-time limits. This Monte Carlo model transports energy without optical phase, so it produces no interference pattern. The mc-photon 0.1.1 manifest lists 11 reference cases as conforming on browser and native GPU hosts. Those results do not cover this slab, its probes or the scattering changes offered here.·simulated · illustrated

Green-and-cyan segmented chains cross a dark field among rounded purple clusters with cyan centres, with smaller gold and orange particles scattered between them.

·Sim Lab·in development·simulated · illustrated

Segmented chains: a particle world found by automated search

8,192 simulated particles of five kinds form green-and-cyan chains and purple clusters with cyan centres. Their attraction table was found by an automated search and scored by open vision models.

Film seconds 20–35 of 60, one starting seed and a drifting camera: green-and-cyan chains glide among purple clusters while smaller particles stream between them.

Simulated patterns, not biological life. Vision-model scores do not show that people find them more lifelike. Glow, trails, colours and camera are illustrated.

Methods

Particle Life: 8,192 particles of five kinds follow an asymmetric attraction table and Tom Mohr's force law in the Sim Lab WebGPU engine. Automated search selected table s1-0376 using SigLIP SO400M-14-384 through open_clip; CLIP ViT-B/32 was also scored. The full 60-second film uses starting seed 237587 and model steps 240–3,840. Glow, trails, colours and camera are illustrated.·simulated · illustrated

Six panels: the bowl's pressure field through the template skull and brain from the published reference, our solver and k-Wave; each code's difference from the reference; and slab errors along the beam of about 10–16 % for both sources and codes.

·Brain·in development

Brain inside and 200 periods: this solver and k-Wave both pass benchmark 8

With brain inside the template skull, our solver's whole-field L2 error against the published k-Wave reference is 11.95 % (bowl) and 12.63 % (piston), under the case limits of 29 and 28 %.

Bowl, y = 0: the published reference computed at 0.167 mm, then our solver and k-Wave at 0.5 mm (top); each code minus the reference, and L2 error in 20 mm slabs for both sources and codes (bottom).

Exploratory: the brain interior and longer read followed earlier failures. Between codes, L2 is 5.1 % (bowl) and 4.1 % (piston), using k-Wave as the reference.

Methods

Brain Atlas WebGPU FDTD 0.3.1 and k-Wave via k-wave-python 0.6.2 (CUDA): 0.5 mm, 200 periods, brain inside the MNI-template skull; ITRUSST benchmark 8, bowl and piston, 500 kHz. Runs read over periods 193–200; published reference computed at 0.167 mm, read at period 200. Metrics: our reimplementation of the intercomparison library on the 0.5 mm volume. Aubry et al. 2022, ITRUSST phase 1, Zenodo 6020543, CC BY 4.0.·simulated

Inside the simulated rat neuromuscular junction at 0.600 ms: thousands of acetylcholines, each coloured by its fate at 5 ms, spread from the release site across the pink folds of the muscle membrane.

·Cells·in development

Where 6,200 acetylcholines go: each coloured by its fate at 5 ms

One simulated release in MCell's rat junction model: 6,200 acetylcholines coloured by their fate at 5 ms; 1,530 split by enzyme, 3,924 lost at the model's rim, 746 still in the cleft.

The stream's 12 s cover 0.25–1.75 ms of the run, as the release spreads through the folds. The full film ends on the tally. The rim is the simulated patch's edge, not a fate in the body.

Methods

MCell's rat neuromuscular junction model (MCell test suite, MIT licence) at its own setting, run by Brain Atlas's WebGPU kernel on native Metal, recording all 6,200 acetylcholines every 1 µs for 5 ms. The engine's counts equal an independent count of the final recorded states. The rim is the simulated patch's edge, where molecules leave the model. Camera, look and colours are artist-made; positions are recorded, joined by straight lines.·simulated · illustratedMCell test suite © 2006–2021 Salk Institute for Biological Studies and Pittsburgh Supercomputing Center, Carnegie Mellon University (MIT licence).

A dark blue optics bench: on the left a sharp spoke-and-bar test chart, on the right the same chart as a soft blur, above a ray diagram through an eye and small plots of point spread and contrast.

·Sim Lab·in development·simulated · illustrated

Wave Bench: bring light into focus in a simulated eye

A new Sim Lab bench: change the target, pupil and focus, then compare the source with the simulated retinal image and a brightness profile across it.

Film seconds 0–15 of 30: a chart seen through a schematic eye; add nearsightedness and the simulated retinal image blurs; add matching spectacles and it sharpens again.

A schematic eye, not a real one; an earlier version failed its comparison with published measurements of central vision's contrast transfer.

Methods

The authored chart, sinusoid or pair of finite Gaussian spots is viewed through a centred schematic eye at 555 nm. Conic corneal and lens surfaces and a gradient-index lens form the rays and point-spread function; an ideal phase screen represents spectacles. Image convolution uses that point spread over one 4° viewing plane. The lower probe reads linear intensities from the source and retinal textures, before display gamma. Its finite-grating contrast fit is separate from the optical modulation transfer function. Automatic accommodation selects a local focus with an authored lens deformation. The ray, point-spread and modulation-transfer calculations were checked against double precision at ten sampled settings on Apple and NVIDIA GPUs; those checks exclude source rendering, image convolution and autofocus, and are not a comparison with real eyes. An earlier version failed its comparison with published measurements of central vision's contrast transfer (EYE-001). Target geometry, colour and chapter timing are illustrated.·simulated · illustrated

A dark field crossed by winding filaments of alternating orange and yellow segments, looping around round cyan nodes scattered between them.

·Sim Lab·in development·simulated · illustrated

Primordial soup: a filament network found by automated search

Two Particle Life pages: an initial world, and a filament network whose particle rules were found by automated search and scored by open vision models.

Film seconds 0–15 of 30: orange-and-yellow filaments wind round cyan nodes. Each colour is a kind of particle following one table of attractions and repulsions.

A simulation, not biological life; a score is a vision model's judgement of a picture. Glow, trails, colours and camera are illustrated.

Methods

Particle Life: 8,192 particles of 3 kinds under one attraction table found by an automated search scored with two open vision models through open_clip (SigLIP as the criterion, CLIP also scored), with Tom Mohr's force law, run in WebGPU on the Sim Lab's neighbour-bin engine.·simulated · illustrated

Two modes, one engine.

One viewport, one way of working, one Methods line under every image. Each mode is a part of the body with its own scenes.

The whole brain Every bundle in the atlas 20,000 tracers ride the heartbeat Arteries and veins on the living cortex LSD seated in its receptor One neuron

Brain · live

A living human brain you can fly through, from the whole head to a single synapse.

Real anatomy, blood flow and stroke, a seizure spreading under its electrodes, neurons firing on their own arbors, drugs at their receptors, a coil over the motor cortex, a cortex taking shape, and the visual world on V1. Nine scenes, every one running in your browser.

Methods
The whole brain
BigBrain 2015's own pial surface (Amunts et al. 2013), 163,842 vertices a hemisphere, carried into MNI152 2009c through its recorded template warp. Native Metal still at 3840×2160, 2.0 s into the 18 s dive (stage 1 of 8, hd-cortex, 0–4 s), engine 3fd56e63, the 2026-10-03 powers-4k-v1 capture set with SHA256SUMS. · real data
Every bundle in the atlas
HCP1065 (Yeh 2022), 91 bundles simplified to 15.5 M points at 0.1 mm · wgpu engine, 2940×1912 · real data
20,000 tracers ride the heartbeat
3D lattice-Boltzmann CFD at 0.1 mm on IXI165's TOF-MRA lumen (242,682-vertex wall), inlets from the 1D v4 model, cycle periodic to 0.3 % · Blender 5.2 Cycles, 1080p at 4.17× slower than life · simulated
Arteries and veins on the living cortex
BigBrain pial surface · Mouches & Forkert arterial atlas and VENAT 7 T venous atlas · the circle of Willis from BodyParts3D, © The Database Center for Life Science (DBCLS), CC BY-SA 2.1 Japan, built on the TARO model from one volunteer's MRI (Nagaoka et al. 2004); reworked by Z-Anatomy (Gauthier Kervyn), CC BY-SA 4.0 · CoW 1D v0.1 pulse · HCP1065 bundles · Blender 5.2 Cycles, 1024 samples, 2560×1440. This image is shared under the same licence, CC BY-SA 4.0. · real data · simulated · illustrated
LSD seated in its receptor
PDB 6WGT (LSD-bound 5-HT2A) morphing to 9WPQ, Molecular Nodes; occupancy from the PK course, 96 min after 200 µg · Blender 5.2 Cycles, frame 241, 512 samples, 2560×1440 · real data · simulated · illustrated
One neuron
One of H01's 104 proofread neurons, with its synapses (Shapson-Coe et al. 2024). Native Metal still at 3840×2160, 18.0 s into the 18 s dive (stage 8 of 8, h01-proofread-neuron, 15–18 s), engine 3fd56e63, the 2026-10-03 powers-4k-v1 capture set with SHA256SUMS. · real data
The chest of the Visible Human Male under one light: shoulders, ribs and lungs, the great vessels rising between them The whole skeleton, with the muscles of the left half laid over it

Body · preview

A whole body on real CT anatomy, with system switches and structures named by a click.

Explore the whole body, then take the dive: from the heart into a mouse heart muscle cell's myofibril and sarcomere, imaged by electron tomography, and on to the cardiac thick and thin filaments from published structures. Watch the heart's simulated beat, and the Walk, a recorded walking trial replayed on a musculoskeletal model. Next come the circulation out to the brain, then the lungs.

Methods
The chest of the Visible Human Male under one light: shoulders, ribs and lungs, the great vessels rising between them
Cycles render of the Body pack (body-assets-vhp-male 0.4.4): Visible Human Male (NLM), one donor; head removed; segmented with TotalSegmentator (Wasserthal et al. 2023); elbows (no CT there) are a network's prediction from cryosection colour, edited by rules set after seeing its renders; the right humerus end and left forearm top aren't shown. HRA v2.5 organ parts and heart (HuBMAP), placed by a rigid torso/pelvis fit · real data · simulated · illustrated
The whole skeleton, with the muscles of the left half laid over it
Whole-body model: BodyParts3D, © The Database Center for Life Science (DBCLS), CC BY-SA 2.1 Japan, built on the TARO model from one volunteer's MRI (Nagaoka et al. 2004); reworked by Z-Anatomy (Gauthier Kervyn), CC BY-SA 4.0; body-zanatomy 0.1.5; face removed by a shear-plane cut; Blender Cycles render. This image is shared under the same licence, CC BY-SA 4.0. · illustrated

What's being built.

Two showpieces still to land, each on real data at every scale it crosses. They join Body as they land, each with its Methods line.

Every image names its data.

Every image names its model and its data in a Methods line and ends it with what you're looking at: real data, simulated or illustrated. When a model misses the measured data, its card says so and stays up. Faces are removed from every scan.