⚛️ quantum physics

Formulation and evaluation of ocean dynamics problems as optimization problems for quantum annealing machines

This paper demonstrates that while simulated annealing successfully solves the classical Stommel ocean dynamics problem by casting it as an optimization task, current quantum annealing hardware faces significant limitations due to restricted connectivity, highlighting the need for hardware and algorithmic improvements before quantum machines can effectively model geophysical dynamics.

Original authors: Takuro Matsuta, Ryo Furue

Published 2026-03-24

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Takuro Matsuta, Ryo Furue

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to predict the weather or the movement of ocean currents. These are incredibly complex puzzles. The ocean and atmosphere are like giant, swirling fluids where every drop interacts with every other drop. To solve these puzzles, scientists usually use supercomputers to simulate the physics. But as our computers get older and hit a wall (Moore's Law is slowing down), we need a new kind of engine to keep solving these problems.

Enter Quantum Annealing. Think of this paper as a "test drive" report for a new, futuristic car engine trying to solve an old, classic navigation problem.

Here is the breakdown of what the researchers did, using simple analogies:

1. The Problem: Turning a River into a Puzzle

The scientists wanted to solve a famous, simplified model of ocean currents called the Stommel Problem.

The Old Way: Usually, you solve this by breaking the ocean into a grid (like a chessboard) and doing math on every square. It's like trying to walk through a forest by stepping on every single leaf.
The New Way (The Paper's Idea): Instead of walking step-by-step, they turned the whole problem into a minimization game. Imagine you have a bumpy landscape (a mountain range with valleys). The goal is to find the deepest valley (the lowest point). In physics, the "depth" of the valley represents the "error" in your solution. The deeper the valley, the better the answer.

2. The Two Hikers: Simulated Annealing (SA) vs. Quantum Annealing (QA)

To find that deepest valley, the researchers used two different "hikers" (algorithms):

Simulated Annealing (SA) – The "Thermal Hiker":
Imagine a hiker trying to find the bottom of a valley. If they hit a small bump, they might get stuck. To escape, this hiker is allowed to jump randomly. Sometimes they jump up a hill (which seems silly), but this helps them escape small, shallow pits so they can find the deepest valley. This is like shaking a box of marbles to settle them into the lowest spot.
- Result: This method worked perfectly. It found the correct ocean current patterns every time.
Quantum Annealing (QA) – The "Ghost Hiker":
This is the fancy new tech. Instead of just walking or jumping, imagine a "ghost" hiker who can be in many places at once (quantum superposition). They can also "tunnel" through walls (quantum tunneling) instead of climbing over them.
- The Catch: The researchers used a real machine made by D-Wave. Think of this machine as a very specific, rigid maze. The "Ghost Hiker" can only move between specific connected rooms. If the problem requires the hiker to jump between two rooms that aren't connected, the machine has to build a long, complicated bridge (called graph embedding) to make it work.
- Result: The Ghost Hiker got lost. Because the real ocean problem is too complex and the machine's "rooms" are too limited, the hiker couldn't build the right bridges. The machine got stuck in shallow pits and gave wrong answers, especially when the problem got bigger.

3. The Two Maps: Grid vs. Waves

The researchers tried two ways to draw their map of the ocean:

The Grid (Finite Difference): Breaking the ocean into a square checkerboard.
The Wave (Spectral Expansion): Describing the ocean as a mix of smooth waves (like a musical chord).

The Grid: The "Thermal Hiker" (SA) did great here. The "Ghost Hiker" (QA) failed unless the map was tiny (low resolution).
The Wave: The "Ghost Hiker" did slightly better here, but still struggled. The problem is that in the "Wave" method, every note interacts with every other note. This creates a massive, tangled web of connections that the rigid D-Wave machine simply cannot handle with its current wiring.

4. The "Zoom" Trick

To make the math work on these machines, the researchers used a clever trick called iteration.

Imagine trying to guess a number between 1 and 100.
Round 1: You guess "50".
Round 2: You realize it's between 40 and 60, so you "zoom in" and guess "45".
Round 3: You zoom in again.
The researchers found that doing this "zooming" with a small number of variables worked better for the Quantum Machine than trying to guess the whole number in one giant leap with thousands of variables.

The Bottom Line: What Does This Mean?

The Good News: The math works. We can turn ocean and weather problems into optimization games that quantum computers can theoretically solve. The "Thermal Hiker" (classical simulation) proved the concept is sound.
The Bad News: The current Quantum Machines (like the D-Wave) are too small and too rigid. They are like a toy car trying to drive a Formula 1 race. They get stuck because the "roads" (connectivity) on the chip aren't connected enough to handle the complex traffic of ocean dynamics.
The Future: If we build bigger quantum computers with more connections (better "roads") and better software to map the problems onto them, this technology could revolutionize how we predict climate change and ocean currents.

In a nutshell: The researchers successfully taught a classical computer how to solve an ocean puzzle using quantum logic. They then tried to hand the puzzle to a real quantum computer, but the computer was too small and had too few connections to solve it. It's a promising start, but the hardware needs a serious upgrade before it can replace our supercomputers.

1. Problem Statement

The paper addresses the challenge of solving geophysical fluid dynamics problems (specifically ocean and atmospheric circulation) using Quantum Annealing (QA).

Context: Traditional High-Performance Computing (HPC) faces limitations due to the end of Moore's Law and the difficulty of parameterizing subgrid-scale interactions in climate models.
Goal: To determine if QA machines (specifically D-Wave systems) can solve partial differential equations (PDEs) describing ocean dynamics by reformulating them as combinatorial optimization problems.
Core Challenge: QA machines operate on binary spin variables ( $\sigma \in \{-1, 1\}$ ) and have limited hardware connectivity, whereas ocean dynamics are described by continuous real-valued variables on complex grids.

2. Methodology

The authors propose a framework to transform linear PDEs into Ising models (quadratic unconstrained binary optimization problems) suitable for annealing.

A. Problem Formulation

The general linear PDE is defined as $L[f](\mathbf{x}) = 0$ . The solution is found by minimizing the $L_2$ norm cost function:
$H = \|L[f]\|^2 = \int_0^1 \int_0^1 |L[f](x,y)|^2 dx dy$

B. Discretization Strategies

Two methods are employed to discretize the cost function $H$ :

Finite Difference Method: The domain is gridded ( $N \times N$ ). The PDE becomes a system of linear equations $A\mathbf{w} - \mathbf{v} = 0$ . The cost function is minimized via least squares, resulting in a quadratic form $H(\mathbf{w}) = \sum J_{ij}w_i w_j + \sum h_i w_i$ .
Truncated Spectral Expansion: The solution $f(\mathbf{x})$ is approximated by a truncated series of orthogonal basis functions (e.g., Fourier sine series). The cost function is expressed in terms of expansion coefficients.

C. Mapping to Ising Model

Since QA machines require binary inputs, real-valued coefficients ( $w_i$ ) are mapped to spin variables ( $\sigma$ ) using a binary expansion:
$w_i = c_i + s_i \sum_{\alpha=0}^{n_{spin}} \frac{\sigma_i^{(\alpha)}}{2^\alpha}$

Iteration Procedure: To improve precision without exploding the qubit count, the authors use an iterative "zooming" approach. In each epoch, the solution is refined by narrowing the search range around the previous result.

D. Annealing Algorithms

The study compares two solvers:

Simulated Annealing (SA): A classical probabilistic algorithm using thermal fluctuations (Metropolis algorithm) to escape local minima.
Quantum Annealing (QA): Uses a D-Wave Advantage_system4.1 machine. It utilizes quantum tunneling and transverse fields to find the ground state.
- Constraint: Due to the D-Wave's limited connectivity (Pegasus graph), graph embedding is required. This maps logical spins to physical qubits, often requiring multiple physical qubits to represent one logical spin and adding penalty terms to the Hamiltonian.

3. Key Contributions

Framework Development: Successfully formulated the classical Stommel problem (a wind-driven ocean gyre model) as an optimization problem solvable by QA.
Method Comparison: Systematically compared Finite Difference vs. Spectral Expansion, and Iterative vs. Direct (large spin number) binary encoding.
Hardware Evaluation: Provided a realistic assessment of current D-Wave hardware capabilities and limitations for geophysical applications.
Nonlinear Extension: Demonstrated the theoretical extension of the spectral expansion method to a simple nonlinear ordinary differential equation (requiring degree reduction algorithms to handle 4-spin interactions).

4. Results

A. Simulated Annealing (SA) Performance

Success: SA successfully reproduced the "true" solution (calculated via standard linear algebra) for the Stommel problem.
Hyperparameters: Accuracy was highly sensitive to the scale factor ( $S$ $S$ ) and number of spins ( $n_{spin}$ $n_{s p in}$ ).
- A scale factor of $S=0.8$ yielded better results than $S=0.5$ .
- Increasing $n_{spin}$ from 2 to 3 improved accuracy, but further increases yielded diminishing returns compared to optimizing $S$ .
Conclusion: SA is a robust classical baseline for this approach.

B. Quantum Annealing (QA) Performance

High Resolution Failure: For the $11 \times 11$ $11 \times 11$ grid (Finite Difference), QA failed to find good solutions. The cost function remained high, and the streamfunction was incorrect.
- Causes: Hardware noise, excessive annealing time requirements, and the overhead of graph embedding (which introduced too many penalty terms and qubits).
Low Resolution Success: When the grid was reduced to $5 \times 5$ , QA successfully reproduced the solution with a cost function $< 10^{-2}$ .
Spectral Expansion:
- For low-order expansions ( $n_x=10$ ), QA matched SA results.
- For higher orders ( $n_x=20$ ), QA failed to converge to a reasonable gyre structure. The dense connectivity of spectral modes made graph embedding inefficient, leading to a proliferation of auxiliary qubits and errors.
Nonlinear Test: In a simple nonlinear ODE test, both SA and QA found the correct coefficients, but QA had lower accuracy (Cost $\sim 10^{-4}$ ) compared to SA (Cost $\sim 10^{-7}$ ).

C. Iteration vs. Large Spin Number

Iteration: Using an iterative procedure with few spins per variable ( $n_{spin}=2$ ) allowed SA to converge to high accuracy.
Direct Encoding: Using a large number of spins ( $n_{spin}=5$ or $8$) in a single shot worked well for SA but failed for QA, even with long annealing times ( $1000 \mu s$ ). The increased qubit count exacerbated noise and embedding issues.

5. Significance and Conclusion

Hardware Limitations: The primary bottleneck for applying QA to ocean dynamics is the limited connectivity of current quantum annealers. Graph embedding significantly increases the problem size and introduces errors, making high-resolution or spectral methods currently impractical.
Potential of SA: Simulated Annealing proves that the formulation (casting PDEs as optimization problems) is valid and effective. It suggests that if hardware improves, the methodology is sound.
Future Outlook:
- Practical application requires either hardware improvements (higher connectivity graphs) or better graph embedding algorithms.
- While current QA machines are not yet faster or more accurate than classical solvers for these specific problems, the potential for speedup in the future (polynomial vs. exponential time) remains a motivating factor.
- The paper concludes that while quantum gate computing and tensor networks are also promising, QA offers a distinct path for combinatorial optimization in geophysics, provided the hardware constraints are overcome.

In summary, the paper demonstrates that while the mathematical formulation of ocean dynamics for quantum annealing is feasible, current hardware limitations (connectivity and noise) prevent the solution of realistic, high-resolution problems. The approach is currently viable only for coarse-grained or low-order spectral approximations.