Most of what we know about how the brain flushes out waste, including the amyloid-beta protein that builds up in Alzheimer's disease, comes from mice. In mice, cerebrospinal fluid (CSF, the clear fluid that bathes the brain) is pushed through narrow channels around blood vessels by the rhythmic pulsing of arteries. This proposal asks whether a second, slower force is missing from that picture in humans: buoyancy. The brain's core runs slightly warmer than its surface, and warm fluid rises. Engineers model this with the Boussinesq approximation, which treats fluid density as constant except in the gravity term, so that small temperature differences drive slow circulating currents.
Buoyancy is irrelevant at the scale of a mouse brain. The quantity that decides whether it matters, the Rayleigh number (buoyant driving divided by viscous and thermal damping), grows with the cube of the fluid compartment's size. A human brain's CSF-filled ventricles and surface spaces are roughly ten to fifteen times larger than a mouse's, which makes the Rayleigh number about a thousand times larger. There is a second reason buoyancy could punch above its weight. Arterial pulsation is oscillatory, so most of its fluid motion goes back and forth, and the net one-way transport it produces is small. A weak but steady buoyant drift therefore does not have to beat the pulsation amplitude. It only has to rival the small net flow that pulsation leaves behind.
The core claim is testable. For physiological temperature gradients, human ventricular and subarachnoid CSF may sit in a mixed-convection regime, where buoyancy contributes a measurable share of net solute transport. Mouse perivascular spaces sit firmly in the regime where pulsation dominates. If this holds, it gives a concrete, physics-based reason why the mouse clearance results of Multisensory gamma stimulation promotes glymphatic clearance of amyloid may not transfer to humans. It also predicts that the head's orientation relative to gravity changes net CSF drift.
Core Claim and Why It Is New
Murdock et al. showed that 40 Hz sensory stimulation raises arterial pulsation power near 0.1 Hz and drives CSF influx and amyloid removal in 5XFAD mice. Blocking the astrocyte water channel aquaporin-4 abolished the effect, which establishes a pulsation-and-water-channel mechanism. They treat the fluid as tracer movement and explicitly flag the mouse-to-human scale gap as unresolved. This work motivates the proposal: it identifies the gap but does not formulate transport equations.
Early role of vascular dysregulation on late-onset Alzheimer's disease used 1,171 subjects from the Alzheimer's Disease Neuroimaging Initiative (ADNI) to order disease events. It placed vascular dysregulation first, before amyloid deposition. The model is statistical and has no fluid mechanics. In our framing, weakened vascular forcing lowers the pulsation-driven term, so the ratio of buoyancy to forcing rises. Buoyancy would then matter more exactly when vascular function declines.
No paper in the file computes this regime map for brain CSF, and none asks whether it differs between species. The contribution is a dimensionless scaling audit, a set of geometry-resolved mixed-convection simulations, and a posture experiment that could falsify the idea.
The governing model is the Boussinesq momentum equation used in Entropy Generation Rates in Two-Dimensional Rayleigh–Taylor Turbulence Mixing, coupled to a transported solute. The regime is summarized by a Richardson number adapted to oscillatory forcing:
$$Ri_{net} = \frac{g\,\beta\,\Delta T\,L}{U_{net}^2}$$
Here $g$ is gravity, $\beta$ is the thermal expansion coefficient of CSF, $\Delta T$ is the core-to-surface temperature difference across a compartment of size $L$, and $U_{net}$ is the time-averaged net velocity left over from pulsation. This differs from the classical lid-driven definition, which uses the forcing velocity itself, as in A Numerical Study of Mixed Convection in Square Lid-Driven with Internal Elliptic Body. That study found that $Ri \ll 1$ gives smooth, forcing-dominated fields and $Ri \gg 1$ gives recirculating buoyant cells. Using $U_{net}$ instead of the oscillation amplitude is the key conceptual step.
Research Program
The first stage is a back-of-envelope regime map across mouse perivascular spaces, human perivascular spaces, human ventricles, and the human subarachnoid space. It sweeps $\Delta T$ from 0.1 to 1 K and uses ranges of $U_{net}$, with uncertainty carried through explicitly. As a rough illustration, water at 0.1 K over 100 µm gives a Rayleigh number of order $10^{-3}$, so buoyancy in perivascular channels is almost certainly negligible. At centimeter scale the same gradient reaches order $10^3$, which is near the threshold for convection. This stage decides whether the rest is worth doing.
The second stage runs geometry-resolved simulations with a standard finite-element solver. The domains are human ventricles and subarachnoid spaces segmented from Human Connectome Project anatomical scans, and mouse brain geometry for comparison. The simulations impose oscillatory wall motion at cardiac and 0.1 Hz vasomotion frequencies and a thermal boundary condition. The output is net tracer transport with the buoyancy term switched on and off.
For the smaller perivascular geometry, we use channels segmented from the MICCAI 2024 dataset of enlarged perivascular spaces (the fluid-filled channels around small vessels, visible on MRI when enlarged) (Standardized Evaluation of Automatic Methods for Perivascular Spaces Segmentation). The expected result there is that buoyancy is negligible, which serves as an internal check.
We also check where the Boussinesq assumption itself breaks down, using temperature-dependent viscosity. The file flags non-Boussinesq effects as the approximation's main failure mode.
The third stage is inverse inference. The method of Hybrid Boundary Physics-Informed Neural Networks separates boundary fitting from interior physics in tortuous domains. We use it to fit sparse velocity measurements while estimating the buoyancy coefficient. It is a tool for sparse data, not the primary forward solver.
Baselines
The naive reference point is a pulsation-only model: the identical solver and geometry with the buoyancy term removed. Every claim about buoyancy is stated as the difference in net tracer transport relative to this model. The two runs share mesh, time step and forcing parameters, so the gap measures the physics term and not a solver artifact. Buoyancy "matters" if it changes net transport by more than the uncertainty band produced by sweeping forcing amplitude within physiological ranges.
The strongest existing competitor for explaining clearance is the pulsation-and-water-channel mechanism of Murdock et al. It is supported by causal perturbations in mice, and no fluid simulation is reported on Lacuna for it. We implement it as the pulsation-only model with a permeable wall standing in for water-channel exchange. It is beaten only if adding buoyancy improves the fit to posture-dependent human flow data, described below.
For the inverse-inference step, the strongest method baseline is the hybrid boundary network. It reports a mean squared error of 0.00088, versus 0.00801 for a standard physics-informed network, on the obstructed-cavity benchmark. We first reproduce roughly that gap on that benchmark before trusting it on brain geometry.
For the imaging-based comparison, the existing clearance proxy is the ALPS diffusion index. The TBI study reports that it predicts plasma neurofilament light with b = −0.482, p = 0.039 (Glymphatic clearance estimated using DTI-ALPS is reduced after traumatic brain injury). No reported result was found on Lacuna for a mechanistic transport model predicting any Alzheimer's biomarker.
The key ablation varies $L$ with all else held fixed. Buoyancy's share of net transport should grow roughly as $L^3$ in the Rayleigh number, and correspondingly in transport. Shape artifacts would not follow that law.
Why not just…?
Why not just compute the Rayleigh number and stop? For perivascular channels we largely will, because the estimate already says buoyancy is negligible there. For centimeter-scale compartments, however, the classical Rayleigh or Richardson estimate compares buoyancy to the pulsation amplitude. The relevant comparison is to the much smaller net flow, which only a simulation of oscillatory forcing in real geometry gives. If the simulated $Ri_{net}$ stays below about 0.01 in every human compartment across plausible $\Delta T$, we report a clean negative result. That result would be useful in its own right, because it lets future clearance models drop buoyancy with justification.
Why not just use the ALPS diffusion index as the human measure of clearance? White Matter Geometry Confounds DTI-ALPS Measures showed that the index's asymmetry appears across 35 white-matter regions, persists at high diffusion weighting, and tracks axonal geometry. Medullary veins deviate up to 30 degrees from the assumed axis. The index cannot separate fluid drift from tissue structure, let alone a buoyant component, so we use it only as a comparison point.
Why not just extend the data-driven progression model with more biomarkers? The Iturria-Medina analysis orders events but cannot distinguish transport mechanisms. It has no term that responds to gravity or temperature. The posture test below is a direct physical perturbation that no statistical ordering model predicts.
Why not just use standard CFD instead of neural solvers? We agree, and forward simulations use finite elements. The neural method is used only where velocity data are sparse and a parameter must be inferred. If it fails to beat a simple regularized least-squares fit there, we drop it.
Decisive Checks
The first check is the mouse negative control. Simulations at mouse scale must show a buoyancy contribution well below forcing uncertainty. That result would be consistent with Murdock et al.'s finding that blocking water channels abolishes clearance. If mouse simulations instead show large buoyant transport, the model is miscalibrated.
The second, and decisive, check concerns posture. Buoyant drift depends on the angle between the thermal gradient and gravity, but pulsation does not. The model therefore predicts that the time-averaged ventricular CSF velocity field changes in a specific, computable direction between supine and prone positions. This check requires new phase-contrast MRI acquisitions, since no such data are in the file. A confirmed direction-specific shift, of the size the model predicts and absent in pulsation-only simulations, supports the claim. A null result bounds the buoyancy share from above.
Only if both checks pass do we take the Alzheimer's step. That step tests whether ADNI subjects with lower arterial-spin-labeling blood flow, a proxy for weaker vascular forcing, show amyloid deposition patterns aligned with the model's predicted buoyant drift paths. The comparison is against the pulsation-only model in a held-out cross-validation, using regional amyloid PET as the target.
References
- Multisensory gamma stimulation promotes glymphatic clearance of amyloid
- Early role of vascular dysregulation on late-onset Alzheimer’s disease based on multifactorial data-driven analysis
- Entropy Generation Rates in Two-Dimensional Rayleigh–Taylor Turbulence Mixing
- A Numerical Study of Mixed Convection in Square Lid-Driven with Internal Elliptic Body and Constant Flux Heat Source on the Bottom Wall
- Standardized evaluation of automatic methods for perivascular spaces segmentation in MRI - MICCAI 2024 challenge results
- Hybrid Boundary Physics-Informed Neural Networks for Solving Navier-Stokes Equations with Complex Boundary
- Glymphatic clearance estimated using diffusion tensor imaging along perivascular spaces is reduced after traumatic brain injury and correlates with plasma neurofilament light, a biomarker of injury severity
- White Matter Geometry Confounds Diffusion Tensor Imaging Along Perivascular Space (DTI‐ALPS) Measures