HAMON: Could Optical Light Diffraction Replace Digital Math for Time-Series Forecasting?

A new forecasting architecture called HAMON sidesteps learned digital sequence mixing entirely and instead encodes historical data onto an optical aperture, letting light physically propagate through trainable phase masks to produce a prediction. On several standard time-series benchmarks, it outperforms strong digital baselines, cutting mean squared error by up to 14 percent. For embedded and FPGA engineers thinking about the future of edge inference hardware, this is worth understanding now.

What Is the Core Finding?

HAMON shows that a passive optical propagation pass, with no digital temporal mixing layer at inference time, can match or beat learned digital models on certain forecasting tasks. The forecasting computation happens in the physics of light itself.

The motivation comes from an observation that has been building in the research community: simple linear and frequency-domain models are surprisingly hard to beat on long-horizon time-series forecasting. If the underlying operator is approximately linear and low-complexity, the researchers asked whether it even needs to be implemented in learned digital weights at all. HAMON is their answer: encode the problem into an optical field and let diffraction do the work.

How Does It Work Technically?

Diffractive optical computing uses the wave nature of light to perform mathematical operations. In HAMON, historical time-series values are encoded onto an optical aperture as phase or amplitude information. Future time positions are left dark. The light then propagates through cascaded trainable phase masks separated by free-space diffraction regions, and the output field at the detector plane carries the forecast values.

The phase masks are the only learned parameters and they are trained offline using Fourier optics simulation, specifically a framework called TorchOptics. At inference, no matrix multiplication or learned digital mixing occurs. The prediction is read from the intensity pattern of the output field. The paper includes ablations where phase encoding is scrambled or the optical path is bypassed, and those ablations confirm the forecasts are genuinely coming from the optical field rather than from any digital head attached to the output.

What Does This Mean for Embedded and FPGA Engineers?

Right now HAMON is simulated in software, so you cannot drop it onto an STM32 or a Xilinx board today. But it defines a concrete hardware target. Passive diffractive optical systems require no power for the core computation at inference time, which is a striking property for edge deployment where power budgets are tight.

For FPGA designers, the relevant comparison is this: an FPGA implementing a time-series forecasting model still burns watts on multiply-accumulate operations at every inference step. An optical core performing the same operation consumes energy only in the photodetector readout and any analog front end. The training stays digital and offline. If optical hardware catches up to where this simulation points, the inference pipeline could look very different from anything running on silicon today. The Fourier optics basis also means the hardware design space is well understood from decades of optical signal processing work.

What Are the Current Limits?

The results are mixed across benchmarks. HAMON outperforms digital baselines on ETTm2 at all horizons and on ETTh2 at most horizons, and it is competitive on the Weather dataset. However, it trails on the Traffic and Electricity datasets, which have high channel counts. The paper attributes this partly to the architecture not yet being optimized for large numbers of parallel channels.

More practically, the entire system is still a simulation. Fabricating real diffractive optical elements with the precision these phase masks require, and building a readout system that handles intensity-compatible encoding reliably, are open engineering problems. Noise, alignment tolerances, and manufacturing variation in physical optics are not yet characterized for this specific forecasting application. The paper positions HAMON as a proof-of-concept target rather than a deployable system.

As optical computing fabrication techniques continue to improve, architectures like HAMON may eventually give embedded designers a genuinely new class of inference substrate to reach for when power and latency constraints push silicon to its limits.