Its denoising backbone is a one-dimensional time-series Temporal-UNet. To guide the reconstruction, the model fuses a calibrated hardware parameter vector with a jitter spectrum estimate through a conditional encoder, and injects this conditional information into each residual block using feature-wise linear modulation (FiLM) to steer the progressive inverse diffusion process. The generative model is anchored to physics at every step of the inverse process.

A data-consistency projection based on a memory polynomial model of the ADC's nonlinear forward behavior is embedded, solved with a Gauss-Newton approximation. The network output is additionally constrained by an STFT-based L1 term and a spectral convergence term, both designed to suppress intermodulation and phase distortion. The authors achieve low-latency implementation through few-step deterministic sampling, progressive distillation, model pruning, and quantization.

The system also provides parallel multi-sample uncertainty estimation and lightweight online fine-tuning, which the paper states ensure domain adaptability and reliability. The evaluation rests on a hybrid dataset combining simulation and experiment, covering single-tone frequency sweeps, two-tone intermodulation tests, multi-tone signals, and modulation signal scenarios. Input frequencies spanned 10 MHz to 450 MHz with normalized amplitudes from 0.1 to 1.0 of full scale.

At full-scale high-frequency input, the uncompensated converter achieved an effective number of bits (ENOB) of only 4.37 bits. After compensation, the improvements were concentrated at high frequencies. In the band above 300 MHz, total harmonic distortion (THD) was suppressed by approximately 73.04 percent and intermodulation distortion (IMD) by approximately 78.15 percent.

Across the tested scenarios, ENOB increased by 3.21 plus or minus 0.16 bits, signal-to-noise and distortion ratio (SNDR) increased by 11.07 plus or minus 0.76 dB, and spurious-free dynamic range (SFDR) increased by 9.57 plus or minus 0.99 dB. The authors state that these experimental results demonstrate high-fidelity signal reconstruction across different frequency bands and complex operating conditions, and present the method as a reliable technical solution for accurate measurement and communication link optimization in high-speed ADC systems. The reported performance figures carry their stated uncertainties and come from the specific hybrid simulation-and-experiment dataset described in the paper.

The study, published as an open-access article in Scientific Reports on 23 October 2026 by Honglai Yan and XiaoChuan Qian, adds a physics-constrained generative modeling approach to the techniques available for nonlinear distortion compensation in high-speed data conversion.

Subject of Research: Physics-constrained diffusion modeling for nonlinear distortion compensation in high-speed analog-to-digital converters

Article Title: Nonlinear distortion compensation for high-speed ADC using a diffusion model

Article References: Yan, H., & Qian, X. (2026). Nonlinear distortion compensation for high-speed ADC using a diffusion model. Scientific Reports. https://doi.org/10.1038/s41598-026-75603-z

Image Credits: AI Generated

DOI: 10.1038/s41598-026-75603-z

Keywords: ADC, diffusion model, nonlinear distortion, ENOB, intermodulation distortion, SNDR, SFDR, clock jitter, signal processing, machine learning, communications, instrumentation