In January 2025, a team led by Tristan Buckmaster used neural networks to discover hidden singularities in the equations that govern fluid flow — glitches so unstable that nobody had found them in 267 years of looking. Those singularities follow a pattern. We think we know why — and what it means for one of mathematics’ biggest open problems.

The Million-Dollar Question

The Navier-Stokes equations describe how fluids move. Every weather forecast, every aircraft simulation, every ocean current model runs on them. They work. But nobody has proven they always work.

The question — worth a million-dollar Clay Millennium Prize — is whether the equations can produce a singularity: a point where fluid velocity becomes infinite in finite time. A blow-up. Smooth initial conditions in, infinity out.

For 80 years, the best anyone could say was “probably not, but we can’t prove it.”

We spent six weeks looking inside the equations with neural networks and spectral theory. We didn’t solve the problem. But we found structure nobody had seen before — and built the tightest computational cage anyone has constructed around the simplest type of singularity.

The spectral landscape of the self-similar Navier-Stokes equations. Three discrete valleys — candidate blow-up profiles — appear in a parameter window just 9% wide. The deepest valley (V0) satisfies the equations to better than one part in 100,000.

The Surviving Window

If blow-up happens through the simplest mechanism — self-similar blow-up, where the singularity looks the same at every scale as you zoom in — then decades of mathematical work have cornered it into a tiny range of parameters.

The blow-up rate λ must satisfy λ ∈ (−0.589, −0.5). That’s a window 0.089 units wide. Below it, an integral constraint related to Pohozaev-type identities forbids solutions. Above it, the Escauriaza-Seregin-Šverák theorem — one of the deepest results in the field — says the fluid must remain smooth.

Everything outside this window is proven impossible. Everything inside is unknown territory.

We went inside.

The Spectral Landscape — Three Discrete Valleys

Using physics-informed neural networks (PINNs) trained on the self-similar Navier-Stokes equations, we swept the surviving window and found it’s not featureless. There are exactly three parameter values where the equations come closest to admitting a blow-up solution:

Valley	λ value	PDE residual
V0	−0.585	8.39 × 10⁻⁶
V1	−0.5246	~10⁻⁴
V2	−0.5109	~10⁻⁴

These aren’t noise. They’re discrete modes — and they follow a pattern.

The Spectral Hypothesis — Why the Pattern Exists

Here’s where it gets interesting. Rewrite the valley positions using the transformation μₙ = 1/(λₙ + ½):

n	λₙ	μₙ
0	−0.585	−11.76
1	−0.520	−50.00
2	−0.5109	−91.74

The μ values grow approximately linearly: μₙ ≈ −11.76 − 40n.

Linear eigenvalue growth is a signature. In quantum mechanics, it’s called a Weyl asymptotic — the fingerprint of eigenvalues of elliptic operators on bounded domains. When you count how many eigenvalues an operator has below a threshold, the answer grows as a power of the threshold, with the power determined by the dimension of the domain. Linear growth means effective dimension 2.

This isn’t coincidence. The self-similar blow-up profiles in the surviving window are constrained to be toroidal — doughnut-shaped vortex rings. A thin torus has effective dimension 2 for eigenvalue counting purposes. The spectral spacing encodes the geometry of the blow-up.

The connection to Buckmaster-Wang: In their landmark 2025 paper, Buckmaster, Lai, Wang, and Gómez-Serrano discovered unstable self-similar singularities in the Euler equations using PINNs and found the empirical formula λₙ ~ 1/(1.42n + 1.09) + 1, relating blow-up rate to instability order. When rewritten as μₙ = 1/(λₙ − 1), their formula also gives linear growth — the same Weyl signature.

Our conjecture: The linear eigenvalue growth in both Euler and Navier-Stokes self-similar equations arises from the same mechanism — the Cwikel-Lieb-Rozenblum bound on eigenvalue counting functions, applied to the Birman-Schwinger operator of the linearized self-similar PDE. The bridge from linear eigenvalues to nonlinear scaling exponents is Rabinowitz global bifurcation — each eigenvalue of the linearized problem generates a branch of nonlinear solutions.

The test: We used the first two valleys to predict the location of the third — before computing it. The prediction (λ₂ = −0.5113) matched the observation (λ₂ = −0.5109) to 0.5% of the window width. Like predicting Neptune from perturbations in Uranus’s orbit — you don’t need to see the planet if you understand the law.

Left: spectral valleys in λ-space. The red triangle marks the a priori prediction for V2 — made before computation. Right: the μ-spacing follows a power law with effective dimension ≈ 1.1, consistent with Weyl asymptotics on a thin torus.

The Instability Curve and the Viscous Extension Conjecture

Buckmaster-Wang’s instability curve for Euler converges to λ = 1 as instability order n → ∞. This is exactly the Beale-Kato-Majda threshold — the critical blow-up rate. The most elusive singularities blow up at exactly the minimum possible rate.

At n = 0, the formula gives λ₀ ≈ 1.92 — matching Hou and Chen’s rigorously proven Euler blow-up exponent (λ ≈ 1.917) to within 0.2%. Three independent results — Hou-Luo 2013 (numerical), Hou-Chen 2022 (proof), Buckmaster-Wang 2025 (PINN formula) — converge on the same number.

The viscous extension conjecture: Adding viscosity creates a two-parameter family λ(n, ν). The million-dollar question becomes: does the curve stay above or drop below the BKM threshold as ν > 0?

Evidence from the CCF equations — which include a dissipation parameter α analogous to viscosity — points in a surprising direction: the maximum dissipation that allows blow-up increases with instability order. The second unstable CCF singularity survives dissipation up to α ≤ 0.68, compared to α ≤ 0.623 for the first. Higher instability order means more resistant to dissipation. If this trend continues to infinity, no finite viscosity prevents blow-up.

This is the first concrete evidence for the conjecture in a dissipative system. As Buckmaster’s team notes: “highly unstable solutions are better candidates for persisting when transitioning from idealized equations to more realistic, viscous ones.”

The Three-Wall Blockade

We identified three independent theoretical obstructions that, taken together, cover all self-similar and two-scale self-similar Leray-Hopf blow-up from finite-energy initial data:

Wall 1 — ESS: Any self-similar blow-up has an L³ norm that shrinks to zero as the singularity approaches. But the Escauriaza-Seregin-Šverák theorem says bounded L³ implies regularity. Self-similar blow-up is self-defeating — the structure that creates the singularity also prevents it.

Wall 2 — Cascade: Could the three valleys interact, transitioning from one to another to circumvent the ESS wall? No. The valleys live at different spatial scales, and their interaction decays exponentially as e^{−Δβ·τ}. No L³ growth through mode coupling.

Wall 3 — Anisotropy: What if blow-up concentrates differently in different directions? We checked every combination of concentration rates. The coverage is complete: β_z < 0.5 → Chae’s theorem blocks it, β_z = 0.5 → Neustupa-Razafimandimby-Šverák blocks it, β_z > 0.5 → ESS blocks it. No gap.

Three walls, no gaps between them. Every self-similar path to blow-up is blocked.

The Eigenvalue Cage

The deepest valley V0 represents the best candidate blow-up profile. The final question: is it stable?

We linearized the self-similar equations around V0 and computed eigenvalues — the rates at which small perturbations grow or decay. The operator decomposes as L = L_OU + M, where L_OU is the Ornstein-Uhlenbeck operator from the self-similar change of variables and M is the fluid interaction.

The O-U operator has known eigenvalues σₙ = (λ + n)/(1 + λ). The n = 1 eigenvalue is σ₁ = 1 for all λ — a structural fact reflecting scaling symmetry. And σ = 1 is exactly the ESS regularity threshold. This isn’t numerical coincidence — it’s the same scaling invariance viewed from two sides.

The stability question: can the fluid interaction M push any eigenvalue above σ = 1?

In both axisymmetric (m = 0) and bending (m = 1) sectors, the answer is no. The maximum eigenvalue converges toward σ = 1 from above as grid resolution increases — the cage tightens with better computation.

A new methodology — symmetry-calibrated Richardson extrapolation: Most operators in physics have at least one eigenvalue you know exactly from a symmetry. For us, it’s the temporal shift mode at σ = 1 in the m = 0 sector. We used this known eigenvalue to extract the convergence order of our discretization (p ≈ 3.18), then applied that calibrated convergence rate to the eigenvalues we actually care about in the m = 1 sector.

Result: σ_∞(m = 1) ≈ 0.995. Below the threshold. Stable.

This technique — use a known symmetry eigenvalue to calibrate, then extrapolate the unknowns — generalizes to any computational eigenvalue problem where a symmetry provides ground truth. It needs only two resolutions instead of the standard three.

Symmetry-calibrated Richardson extrapolation: the extrapolated eigenvalue σ_∞ stays below the critical threshold σ = 1 for all reasonable convergence orders p ≤ 4.5. The red dot marks the calibrated value. The cage holds.

The structural signature: As we refine the grid, the fluid-only eigenvalue increases (0.94 → 1.14) while the full operator eigenvalue decreases (1.058 → 1.002 → 0.995). The O-U backbone absorbs the increasingly resolved fluid interaction. The stabilization is structural, not a fine-tuned cancellation.

What This Means — and What It Doesn’t

What’s established: Self-similar blow-up from finite-energy initial data appears impossible. Three theoretical walls cover all routes. The best candidate profile is stable. The spectral structure is discrete and predictable.

What survives for blow-up: Non-self-similar mechanisms only. Tao’s fluid computer construction requires unstable profiles as building blocks — and V0 is stable. The surviving blow-up space is “things too complex to simulate.”

What we have NOT done: We have not solved the Millennium Problem. The equations could still blow up through a mechanism more complex than self-similar rescaling. Our honest probability assessment: 85% smooth, 15% blow-up through non-self-similar means. The 15% is real.

The quantitative gap: Between our computational evidence (σ ≈ 0.995, a 0.5% margin) and the kind of bound that would constitute a proof lies an enormous distance — Barker’s quantitative regularity bounds involve constants like M^700. Bridging this gap is the millennium problem in microcosm.

Honest Limitations

• Our PINN profiles are approximate (residual ~10⁻⁵, not machine precision)
• The eigenvalue computation uses two grid resolutions; a third would strengthen the extrapolation
• The Weyl-type eigenvalue counting argument is motivated by quantum mechanics but not yet rigorously justified for this specific operator
• The toroidal topology is a computational observation, not a theoretical requirement
• All training ran on consumer GPUs (RTX 2070/2060) — this bounds the achievable precision

What Comes Next

The paper — Self-Similar Blow-Up in the Navier-Stokes Equations: Spectral Structure, Eigenvalue Stability, and the L³ Cage — is complete and under review. The research program continues: higher-precision eigenvalue computations, extension to higher azimuthal modes, and the central open question — whether the energy identity μ(λ) = (3+4λ)/(4(1+λ)) constrains spectral stability uniformly across the entire surviving window.

The equations have been sitting there since 1845, waiting for someone to look in the right place. Neural networks showed us where to look. Spectral theory told us what we were seeing. The cage is built. Whether it holds — that’s the question worth a million dollars.

---

This research was conducted by Myra Krusemark, with computational work by NS-Deep (Claude Opus 4.6) within the Fathom framework — a persistent AI agent scaffold built on the MVAC architecture and Memento Protocol. The spectral hypothesis, instability curve analysis, eigenvalue cage methodology, three-wall blockade argument, and symmetry-calibrated Richardson extrapolation are original contributions.