Multiplicative PINN Framework

What It Solves#

Physics-informed neural networks (PINNs) suffer from gradient conflicts when enforcing physical laws. Standard PINNs use additive penalty terms — this causes the physics loss to fight the data loss, leading to oscillations, slow convergence, and monotonicity violations.

Key Innovation#

Multiplicative Constraint Enforcement replaces additive penalties:

Traditional: Loss = DataLoss + λ × PhysicsLoss  (additive)
            ↓ (causes gradient conflicts)
Arithmetic Manifold: Loss = DataLoss × ConstraintFactor(PhysicsLoss)  (multiplicative)

The Euler Gate: $G(\mathbf{v}) = \prod_{c} (1 - p_c^{-\tau v_c})$

The Exponential Barrier: $B(\mathbf{v}) = \exp(\gamma \|\mathbf{v}\|_2^2)$

Combined: $C(\mathbf{v}) = G(\mathbf{v}) \times B(\mathbf{v})$

Why It Works#

When constraints are satisfied, $G(\mathbf{v}) \to 1$ and $B(\mathbf{v}) \approx 1$ . When violated, $G(\mathbf{v}) \to 0$ and $B(\mathbf{v}) \to \infty$ . The product zeroes out any solution that violates a constraint.

The deeper reason is that multiplicative enforcement changes the geometry of constraint learning.

Standard PINNs use additive penalties:

L_{\text{add}} = L_{\text{data}} +\lambda_1L_{\text{physics},1} +\lambda_2L_{\text{physics},2} +\cdots.

The optimizer therefore sees a vector sum of gradients. If the data gradient and physics gradients point against each other, they can cancel, oscillate, or distort the loss landscape.

The multiplicative loss instead uses

L_{\text{mult}}=L_{\text{data}}\cdot C(v),

\nabla_\theta L_{\text{mult}} = C(v)\nabla_\theta L_{\text{data}} + L_{\text{data}}\nabla_\theta C(v).

The first term preserves the data-gradient direction and rescales it. The second term adds a structured correction toward constraint satisfaction. This is the main geometric shift: constraints are not added as competing forces; they become an observer kernel over the gradient flow.

In the reported experiments, this structure enforced multiple constraints simultaneously, including monotonicity, Lipschitzness, positivity, and convexity. The showcased monotonicity experiment reached $0.00\%$ violations. For Navier-Stokes, the residual improved from approximately $2.8\times10^{-3}$ to $10^{-5}$ , a $99.64\%$ reduction.

The incompressibility result is partly architectural. Using a streamfunction

u=\frac{\partial\psi}{\partial y}, \qquad v=-\frac{\partial\psi}{\partial x},

gives

\nabla\cdot u = \frac{\partial^2\psi}{\partial x\partial y} - \frac{\partial^2\psi}{\partial y\partial x} =0

up to numerical precision, because mixed partials cancel. Thus incompressibility is not merely learned; it is built into the representation.

The performance claim should be read as a benchmark result rather than a universal theorem: the implementation reports about $1{,}000{,}908$ physics-informed states per second and $8000$ time steps in roughly $8$ ms. The mathematical point is independent of the exact benchmark number: multiplicative PINNs preserve the main descent direction while using constraint violations to reshape the measure of the loss landscape.

STOP/Mellin Interpretation#

The Euler gate is not only a prime-weighted penalty. It is a finite reciprocal Euler product.

For a finite set of primes $P$ , define

\zeta_P(s)=\prod_{p\in P}(1-p^{-s})^{-1}.

Then the gate

G(v)=\prod_{p\in P}(1-p^{-\tau v})

G(v)=\frac{1}{\zeta_P(\tau v)}.

So the optimizer does not see the violation coordinate $v$ linearly. It sees it through the Euler-product coordinate

s=\tau v.

The log-gradient of the gate is

\frac{d}{dv}\log G(v) = \tau \sum_{p\in P} \frac{\log p \, p^{-\tau v}}{1-p^{-\tau v}}.

Expanding each geometric factor,

\frac{p^{-\tau v}}{1-p^{-\tau v}} = \sum_{k\ge1}p^{-k\tau v},

gives

\frac{d}{dv}\log G(v) = \tau \sum_{p\in P} \sum_{k\ge1} (\log p)p^{-k\tau v}.

This is the finite-prime analogue of

-\frac{\zeta'(s)}{\zeta(s)} = \sum_{p,k\ge1} \frac{\log p}{p^{ks}} = \sum_{n\ge1} \frac{\Lambda(n)}{n^s}.

Thus the gradient correction from the Euler gate is a truncated von Mangoldt, or prime-power, spectral signal.

In STOP/Mellin language:

\text{constraint violation }v \longmapsto \text{Euler observer coordinate }s=\tau v \longmapsto \text{prime-power spectral response}.

This gives a precise mathematical role to the prime gate:

\boxed{ \text{Euler products define multiplicative observer kernels, and their log-gradients generate prime-power spectral correction fields.} }

The exponential barrier plays a complementary role. The Euler gate is spectral/arithmetic:

G(v)=1/\zeta_P(\tau v),

while the barrier is geometric/convex:

B(v)=e^{\gamma v}.

The gate supplies structured phase-space filtering near the constraint manifold. The barrier supplies survival pressure when violations are large. Clamping the gate near $v=0$ acts as a renormalization floor: it prevents the finite Euler product from annihilating the gradient at the satisfied-constraint boundary.

Superconducting Phase#

At critical $\beta = 1$ , the Riemann zeta function diverges: $\zeta(1) = \infty$

This nucleates a “superconducting phase” where constraints propagate without dissipation — gradients flow freely when physics is satisfied.

Results#

Metric	Additive (Standard)	Multiplicative PINN
Residual reduction (Navier-Stokes)	baseline	99.64%
Monotonicity violations	31.31%	0%
Speedup over CFD	1x	100,000x

Website#

Live: sethuiyer.github.io/multiplicative-pinn-framework

Key Files#

multiplicative-pinn-framework/README.md — Overview
multiplicative-pinn-framework/docs/RESULTS_SUMMARY.md — Detailed benchmarks
DOI: 10.5281/zenodo.18214172

Connection to Core Vision#

The Multiplicative PINN is the proof of concept that multiplicative constraint enforcement works in neural networks. It demonstrates:

The Euler gate structure eliminates gradient conflicts
The Riemann zeta divergence creates a superconducting phase
The approach scales to real PDEs (Navier-Stokes)