Memristor-Based Fourier Neural Operator Delivers 116x Energy Efficiency Gains for Scientific Modeling

Imagine standing in a massive data center, fans roaring, thousands of processors churning away just to simulate a typhoon's path or how a drug molecule diffuses in the human body. Traditional scientific modeling computes every grid point and time step one by one — accurate but painfully slow and power-hungry. A Chinese research team, publishing in *Science Advances*, has unveiled a memristive floating-point Fourier neural operator network that improves scientific modeling energy efficiency by up to 116x and speed by up to 176x, without sacrificing accuracy compared to conventional digital computers.

Why Traditional Scientific Modeling Is Slow and Power-Hungry

Scientific modeling underpins modern research, from climate simulation to materials design. Classical numerical methods solve PDEs point by point on grids — a complex problem can require hours or days on a supercomputer. The core culprit is the von Neumann bottleneck: memory and processing are physically separated, so data shuttles back and forth constantly. In high-precision scientific computing, this involves enormous amounts of floating-point arithmetic, driving both latency and energy bills into the tens of thousands of dollars per simulation.

To escape this, researchers have turned to AI-assisted modeling — specifically the Fourier Neural Operator (FNO).

FNO: The Universal Translator of Physics

The Fourier transform decomposes complex signals into frequency components. FNO learns a mapping operator directly from input functions to solution functions, working efficiently in the frequency domain. Trained even on low-resolution data, it can produce high-accuracy results with zero-shot super-resolution — no iterative stepping required.

For example, the 1D Burgers equation:

$ \partial_t u + \partial_x [u^2/2] = \nu \partial_{xx} u $

models viscous fluid nonlinearity (shocks and turbulence). FNO applies layers of discrete Fourier transforms (DFT) and inverse transforms (IDFT), capturing linear operations in the frequency domain efficiently.

But FNO has a weakness: it requires repeated DFTs, with heavy data movement between processor and memory, and training demands FP32 precision — exactly where traditional digital platforms struggle.

Compute-in-Memory: The Memristor Solution

A memristor remembers the history of current through it, storing data *and* computing on it simultaneously — the warehouse worker who packs shipments in place, rather than hauling goods back and forth.

The team built a heterogeneous CIM system from eight 4-kb memristor chips (32x128 arrays, 1T1R TiN/TaO_x/HfO_2/TiN stack). Two key innovations let imprecise analog arrays reach FP32-level accuracy:

• Exponential pre-alignment and hybrid state mapping (binary + 2-bit hybrid conductance states, average error < 10^-3)
• Fixed parameters (DFT matrices, real/imaginary parts mapped across four arrays, reconstruction error ~5x10^-4, read voltage 0.2V) mapped in-situ; trainable fully-connected layer weights updated on a digital CPU/GPU, avoiding high-precision programming overhead.

During training, Fourier transforms stay on-chip; during inference, all layers migrate onto the memristor chips.

Experimental Results

1D Burgers equation (32-point grid, Gaussian random field initial conditions):

• CIM training loss: 4x10^-3, accuracy 99.6% (digital baseline: 99.8%)
• Inference error: 1.3x10^-2, outputting the t=1s solution directly with no iteration
• Latency: 0.53 ms

3D heat conduction (10x10x10 grid, COMSOL-generated data),

$ \partial u / \partial t = k (\partial^2 u / \partial x^2 + \partial^2 u / \partial y^2 + \partial^2 u / \partial z^2) $

implemented with a single chip via three successive 1D DFTs:

• Training loss: 6.9x10^-3
• Average inference error: 0.24°C, generalizing to unseen heat-source conditions
• Latency: 0.28 ms

| Task | Energy efficiency gain (measured) | Projected at 28nm | Inference latency | Accuracy | |---|---|---|---|---| | 1D Burgers | 20.5x vs GPU | 116x | 0.53 ms | 99.6% | | 3D heat conduction | 5.7x vs GPU | 21x | 0.28 ms | error 0.24°C |

Measured inference energy efficiency reached 3.28 GFLOPS/W (1D) and 23.4 GFLOPS/W (3D). Training energy and latency dropped 16–19%. The 176x speedup comes from millisecond-scale inference versus conventional second-to-day-scale solvers. Practical applications include chip thermal design, where fast 3D temperature prediction helps engineers optimize layouts and prevent overheating.

Outlook

This work extends memristor computing from edge neural networks to high-precision scientific modeling, opening a new chapter for AI-for-Science. Larger chip integrations and algorithm-hardware co-design could tackle turbulence simulation and climate prediction — and eventually bring complex physics simulation to devices far smaller than today's supercomputers.

References

1. Li, J. et al. Memristive floating-point Fourier neural operator network for efficient scientific modeling. *Science Advances*, 11(25), eadv4446 (2025). 2. Li, Z. et al. Fourier neural operator for parametric partial differential equations. *arXiv preprint* arXiv:2010.08895 (2020). 3. Chua, L. Memristor—The missing circuit element. *IEEE Transactions on Circuit Theory*, 18(5), 507-519 (1971). 4. Strukov, D. B. et al. The missing memristor found. *Nature*, 453(7191), 80-83 (2008). 5. Gu, Q. et al. Computing-in-memory paradigm for high-precision scientific computing (2024).