A preprint reports that an open-source framework carried stateless offline machine-learning models into SystemC-AMS simulations with virtually identical metrics to Python baselines after ONNX export. The largest reported gap in R2, one of the scores used to compare predictions with observed values, was 0.003 for an XGB model on the UCI Appliances dataset, and every reported offline-fidelity run finished in under 1.5 seconds. The evaluation used R2, root mean squared error and mean absolute error, along with wall-clock time.
The paper frames SAMpLE around three practical tests: whether Python-model fidelity survives ONNX export, whether heterogeneous models can run through one interface, and whether the system can move between datasets without changes to simulator source. It treats machine-learning models as first-class TDF components and provides native C++ backends for online and offline execution under a unified interface.
Two datasets, time-ordered tests
The data came from two electricity-demand datasets. UCI Appliances contributed 19,735 observations and Tetuan City 52,416; both were sampled every 10 minutes. The experiments split the data in chronological order to avoid temporal leakage, keeping the test design tied to the sequence of observations.
Offline experiments used a chronological 15% test split: 2,954 UCI samples and 7,855 Tetuan samples. Online experiments used a chronological 20% test split: 3,947 UCI samples and 10,484 Tetuan samples. The training, validation and test handling was seed-controlled.
Data preparation was configuration-driven. The pipeline exposes 18 optional keys across five sections and serializes the processed data before creating the seed-controlled, chronology-preserving train, validation and test partition.
Scores and timings were collected on an Intel Core i7-10700 computer with 16 GB of RAM running Ubuntu 22.04. The report gives wall-clock time, R2, root mean squared error and mean absolute error; error values are expressed in watt-hours for UCI and watts for Tetuan.
A common route into simulation
In the interoperability test, three offline backends and eight online backends ran inside the same simulation structure. Switching among the models required no manual code modification, allowing the comparison to keep the surrounding simulation arrangement fixed while the backend changed.
The dataset-portability test changed the input rather than the simulator implementation. Moving from UCI to Tetuan required swapping the input CSV and feature schema, with no changes reported to C++ code, SystemC modules or the build configuration.
At simulation startup, the offline ONNX path checked that the input shape was compatible. It then ran a zero-padded dry run to confirm that the exported graph was callable and returned a finite output.
Online results varied by dataset
The online and offline tests produced different comparisons on UCI. Offline models were reported at R2 values from 0.13 to 0.28, while Hedge Ensemble was the best online method at 0.879. Other online methods occupied a 5-to-30-percentage-point performance band, and the paper describes the top result as more than a threefold improvement over the best offline baseline.
Tetuan produced strong R2 values for both routes. NLMS-ARX reached 0.9961, compared with 0.986 for offline ARX. Its end-to-end inference time was 48 milliseconds; the offline pipeline was reported at 0.6 seconds for Python training plus 0.343 seconds for simulation and ONNX-export overhead.
Runtime varied sharply among the online methods. Across the tests, it spanned three orders of magnitude, from 9 milliseconds for linear filters to 3.6 seconds for Online-GP on Tetuan. These measurements were taken inside the virtual platform under TDF semantics.
The evidence has a defined boundary
The findings are best read as a software-integration evaluation covering regression predictions and runtime on two electricity-demand datasets under the reported SystemC-AMS TDF setup. They do not amount to a universal ranking of machine-learning models, or to evidence about a physical device or real-world electricity demand.
The evidence remains tied to the two named datasets and the reported workstation and simulation setup. The paper states that its configurations, datasets, models, predictions and metric reports were released alongside the source code for exact reproducibility, and describes SAMpLE as open source.
This is an arXiv version 1 preprint dated 26 August 2026. The work received funding from the European Chips Joint Undertaking under Framework Partnership Agreement No. 101139789 (HAL4SDV).
Paper data and sources
Original title: SAMpLE: A SystemC-AMS Machine LEarning-based Framework for Virtual Prototyping
Authors: Andrei Mihai Albu, Sara Vinco
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text