Are Time-Series Foundation Models Ready for E-Nose Data? An Empirical Assessment of Their Embeddings

Mohammed Kamruzzaman; Taeyeong Choi

arxiv: 2606.27672 · v1 · pith:S5JW3DBF · submitted 2026-06-26 · cs.LG

Are Time-Series Foundation Models Ready for E-Nose Data? An Empirical Assessment of Their Embeddings

Taeyeong Choi , Mohammed Kamruzzaman This is my paper

Reviewed by Pith T0 review T1 audit T2 compute T3 formal T4 kernel 2026-06-29 04:51 UTCgrok-4.3pith:S5JW3DBF record.json open to challenge →

classification cs.LG

keywords time-series foundation modelselectronic noseembeddingsgas identificationfine-tuningsensor dataChronosMOMENT

0 comments

The pith

Time-series foundation model embeddings need fine-tuning to work on electronic nose gas data.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper tests whether embeddings from broad time-series foundation models can serve as effective features for identifying gases and estimating their concentrations in electronic nose sensor readings. It evaluates models including Chronos-2 and MOMENT on standard E-Nose datasets and finds that the raw embeddings yield weak results on these tasks. Fine-tuning the foundation models on the gas-sensing data produces usable performance, while further gains occur when the tuned embeddings are combined with features from models built specifically for the prediction tasks. The findings point to both the adaptability of current foundation models to new sensor domains and their need for domain-specific adjustment before reliable use.

Core claim

Embeddings produced by representative time-series foundation models do not directly deliver satisfactory performance for gas identification and concentration prediction on E-Nose data. Fine-tuning the models on the target data is required to reach acceptable accuracy, and fusing the resulting embeddings with representations learned by specialized predictive models yields additional improvement.

What carries the argument

Embeddings extracted from pre-trained time-series foundation models, used as feature inputs for downstream classification and regression on multivariate E-Nose sensor time series.

If this is right

Fine-tuning is required before time-series foundation model embeddings support reliable gas identification or concentration estimates from E-Nose readings.
Fusion of fine-tuned foundation model embeddings with outputs from task-specific models produces higher performance than fine-tuning alone.
Current time-series foundation models show limitations for direct application to gas-sensing data but retain value once adapted.
Hybrid use of foundation-model and specialized representations offers a practical route for deploying these models in sensor domains.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same pattern of needing adaptation may appear when applying current time-series foundation models to other chemical or environmental sensor streams.
Pre-training future foundation models on larger collections of sensor-array data could reduce reliance on per-domain fine-tuning.
The fusion approach could be examined in other time-series domains where off-the-shelf foundation embeddings underperform.

Load-bearing premise

The specific foundation models and E-Nose datasets examined stand in for the full range of time-series foundation models and gas-sensing applications.

What would settle it

Demonstration of a new time-series foundation model that reaches high accuracy on several E-Nose gas-identification and concentration tasks without any fine-tuning or fusion steps.

Figures

Figures reproduced from arXiv: 2606.27672 by Mohammed Kamruzzaman, Taeyeong Choi.

read the original abstract

Inspired by advances in natural language processing and computer vision, "time-series foundation models" (TSFMs) have recently been introduced with the promise of strong generalization across diverse time-series tasks, including forecasting, classification, and anomaly detection, as well as across domains such as healthcare, climate science, and manufacturing. However, their utility for gas-sensing data remains largely unexplored. To address this gap, this paper systematically evaluates recent TSFMs on electronic nose (E-Nose) data. In particular, we investigate whether embeddings produced by representative TSFMs, including Chronos-2 and MOMENT, provide effective representations for gas identification and concentration prediction. Specifically, we show that fine-tuning is necessary to achieve satisfactory performance on E-Nose data, and fusing TSFM embeddings with representations learned by specialized predictive models can further improve the performance, suggesting both the potential and limitations of current TSFMs for gas-sensing applications.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper tests two TSFMs on E-Nose data and finds zero-shot embeddings fall short while fine-tuning plus fusion improves results, but the narrow model coverage and missing experimental details limit how far the conclusions reach.

read the letter

The one or two things to know are that this work applies Chronos-2 and MOMENT to electronic nose data for the first time and reports that their embeddings require fine-tuning to work well for gas identification and concentration prediction, with additional gains from fusing them with specialized models.

The new part is the domain extension to gas sensing, which fills an explicit gap noted in the abstract. The paper does a solid job of running the empirical assessment and pointing out both where the TSFMs fall short in zero-shot settings and where they can be adapted. This kind of targeted check is useful for seeing how well the generalization claims hold up in practice.

Where it gets soft is the reliance on just those two models to speak for time-series foundation models as a whole. If other TSFMs handle the data better out of the box, the necessity of fine-tuning wouldn't apply broadly. The datasets chosen might also share traits like sensor issues or limited scope that aren't representative of all E-Nose setups. The abstract is light on specifics like exact dataset descriptions, metrics used, or comparison baselines, which makes it harder to evaluate how convincing the results are. The stress-test concern about unverified representativeness holds up here.

No formal math or proofs are involved, so that's not an issue. The citation pattern looks standard and relevant to TSFM work.

This paper would interest people in sensor applications or foundation model adaptation for time series. A practitioner in gas sensing or someone exploring new domains for TSFMs could pick up practical insights from the findings.

It deserves a serious referee because it tackles an open question with experiments, even if revisions for more breadth would help.

Recommendation: Send it out for peer review to get feedback on strengthening the experimental design and scope.

Referee Report

2 major / 1 minor

Summary. The manuscript presents an empirical assessment of time-series foundation models (TSFMs) on electronic nose (E-Nose) data. Using embeddings from Chronos-2 and MOMENT, it evaluates their utility for gas identification and concentration prediction tasks. The central claims are that fine-tuning these embeddings is necessary to reach satisfactory performance and that fusing them with representations from specialized predictive models yields further gains, indicating both potential and limitations of current TSFMs for gas-sensing applications.

Significance. If the reported empirical patterns hold under broader testing, the work would document a concrete domain where zero-shot TSFM transfer underperforms and would illustrate a practical hybrid strategy that improves results. This could usefully inform subsequent TSFM development for sensor-derived time series.

major comments (2)

[Abstract] Abstract: The evaluation is restricted to Chronos-2 and MOMENT with the assertion that they are 'representative TSFMs,' yet no justification, selection criteria, or comparison to other models (e.g., additional TSFMs) is supplied. This assumption is load-bearing for the general claim that fine-tuning is necessary across the TSFM class for E-Nose data.
[Abstract] Abstract: The E-Nose datasets are not characterized with respect to sample size, sensor types, gas variety, or potential biases such as drift or small-sample regimes. Without this information it is impossible to determine whether the necessity of fine-tuning and the value of fusion are robust findings or specific to the chosen collections.

minor comments (1)

[Abstract] The abstract states conclusions ('we show that...') without any quantitative metrics, baseline comparisons, or dataset identifiers; even a high-level summary of key numbers would strengthen readability.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on our manuscript. We address each major comment below and outline the revisions we will make to improve clarity and transparency.

read point-by-point responses

Referee: [Abstract] Abstract: The evaluation is restricted to Chronos-2 and MOMENT with the assertion that they are 'representative TSFMs,' yet no justification, selection criteria, or comparison to other models (e.g., additional TSFMs) is supplied. This assumption is load-bearing for the general claim that fine-tuning is necessary across the TSFM class for E-Nose data.

Authors: We agree that the abstract would be strengthened by explicit justification for selecting Chronos-2 and MOMENT. These were chosen as representative due to their recency, public availability of embeddings, and reported strong performance on general time-series benchmarks. In the revised manuscript we will add selection criteria to the abstract and a short paragraph in the introduction, while qualifying that our findings apply to these models and do not claim universality across all TSFMs without further evaluation. revision: yes
Referee: [Abstract] Abstract: The E-Nose datasets are not characterized with respect to sample size, sensor types, gas variety, or potential biases such as drift or small-sample regimes. Without this information it is impossible to determine whether the necessity of fine-tuning and the value of fusion are robust findings or specific to the chosen collections.

Authors: We will add a concise characterization of the E-Nose datasets (sample sizes, sensor types, gas varieties, and notes on drift mitigation) directly into the abstract. The full experimental section already contains these details; elevating a summary to the abstract will make the scope of our empirical findings on fine-tuning and fusion more transparent to readers. revision: yes

Circularity Check

0 steps flagged

No circularity: pure empirical evaluation

full rationale

The paper performs an empirical assessment of TSFM embeddings (Chronos-2, MOMENT) on E-Nose datasets for gas identification and concentration prediction. Claims about the necessity of fine-tuning and benefits of fusion are presented as experimental outcomes, not as derivations, first-principles predictions, or quantities fitted to subsets and then renamed. No self-citations, ansatzes, or uniqueness theorems appear in the abstract or described structure. The work is self-contained against external benchmarks via direct evaluation.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The paper is an empirical assessment relying on standard machine-learning evaluation practices rather than new theoretical constructs.

axioms (1)

domain assumption The selected TSFMs and E-Nose datasets are representative of the broader class and application area.
Invoked when generalizing from specific models and data to statements about TSFM readiness for gas-sensing.

pith-pipeline@v0.9.1-grok · 5693 in / 1118 out tokens · 24307 ms · 2026-06-29T04:51:58.025953+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

11 extracted references · 4 canonical work pages · 1 internal anchor

[1]

arXiv preprint arXiv:2402.03885 , year=

M. Goswami, K. Szafer, A. Choudhry, Y . Cai, S. Li, and A. Dubrawski, “Moment: A family of open time-series foun- dation models,”arXiv preprint arXiv:2402.03885, 2024

work page arXiv 2024
[2]

Chronos-2: From Univariate to Universal Forecasting

A. F. Ansari, O. Shchur, J. K ¨uken, A. Auer, B. Han, P. Mer- cado, S. S. Rangapuram, H. Shen, L. Stella, X. Zhanget al., “Chronos-2: From univariate to universal forecasting,”arXiv preprint arXiv:2510.15821, 2025

work page internal anchor Pith review Pith/arXiv arXiv 2025
[3]

TimeGPT-1

A. Garza, C. Challu, and M. Mergenthaler-Canseco, “Timegpt- 1,”arXiv preprint arXiv:2310.03589, 2023

work page Pith review arXiv 2023
[4]

Benchmarking a time-series foundation model (timegpt) for real-world forecasting applications,

X. Zhang, S. Sridharan, N. H. B. Zahrin, N. Venkataraman, and G. S. Hiong, “Benchmarking a time-series foundation model (timegpt) for real-world forecasting applications,”Ma- chine Learning with Applications, p. 100801, 2025

2025
[5]

Mira: Medical time series foundation model for real-world health data,

H. Li, B. Deng, C. Xu, Z. Feng, V . Schlegel, Y .-H. Huang, Y . Sun, J. Sun, K. Yang, Y . Yuet al., “Mira: Medical time series foundation model for real-world health data,”Advances in Neural Information Processing Systems, vol. 38, pp. 98 657– 98 685, 2026

2026
[6]

How effective are large time series models in hydrology? a study on water level forecasting in everglades,

R. Rangaraj, J. Shi, A. Shirali, R. Paudel, Y . Wu, and G. Narasimhan, “How effective are large time series models in hydrology? a study on water level forecasting in everglades,” arXiv preprint arXiv:2505.01415, 2025

work page arXiv 2025
[7]

From detection to forecasting: utilizing time-series foundation models to anticipate defects in metal additive manufacturing,

J. Zhang, F. Farbiz, M. Jafary-Zadeh, and S. L. Sing, “From detection to forecasting: utilizing time-series foundation models to anticipate defects in metal additive manufacturing,”Journal of Manufacturing Processes, vol. 150, pp. 1040–1052, 2025

2025
[8]

Ultra low power mox sensor reading for natural gas wireless monitoring,

M. Rossi and D. Brunelli, “Ultra low power mox sensor reading for natural gas wireless monitoring,”IEEE Sensors Journal, vol. 14, no. 10, pp. 3433–3441, 2014

2014
[9]

Mox-nw electronic nose for detection of food microbial contamination,

G. Sberveglieri, G. Zambotti, M. Falasconi, E. Gobbi, and V . Sberveglieri, “Mox-nw electronic nose for detection of food microbial contamination,” inSENSORS, 2014 IEEE. IEEE, 2014, pp. 1376–1379

2014
[10]

Calibration transfer and drift counteraction in chemical sensor arrays using direct standardization,

J. Fonollosa, L. Fernandez, A. Guti ´errez-G´alvez, R. Huerta, and S. Marco, “Calibration transfer and drift counteraction in chemical sensor arrays using direct standardization,”Sensors and Actuators B: Chemical, vol. 236, pp. 1044–1053, 2016

2016
[11]

Mdfe-net: A meta-learning driven dual-branch feature extraction network for e-nose sensor drift adaptation,

Q. Yang, J. Liu, Y . Shi, Y . Wang, and H. Men, “Mdfe-net: A meta-learning driven dual-branch feature extraction network for e-nose sensor drift adaptation,”ACS sensors, vol. 11, no. 5, pp. 4057–4067, 2026

2026

[1] [1]

arXiv preprint arXiv:2402.03885 , year=

M. Goswami, K. Szafer, A. Choudhry, Y . Cai, S. Li, and A. Dubrawski, “Moment: A family of open time-series foun- dation models,”arXiv preprint arXiv:2402.03885, 2024

work page arXiv 2024

[2] [2]

Chronos-2: From Univariate to Universal Forecasting

A. F. Ansari, O. Shchur, J. K ¨uken, A. Auer, B. Han, P. Mer- cado, S. S. Rangapuram, H. Shen, L. Stella, X. Zhanget al., “Chronos-2: From univariate to universal forecasting,”arXiv preprint arXiv:2510.15821, 2025

work page internal anchor Pith review Pith/arXiv arXiv 2025

[3] [3]

TimeGPT-1

A. Garza, C. Challu, and M. Mergenthaler-Canseco, “Timegpt- 1,”arXiv preprint arXiv:2310.03589, 2023

work page Pith review arXiv 2023

[4] [4]

Benchmarking a time-series foundation model (timegpt) for real-world forecasting applications,

X. Zhang, S. Sridharan, N. H. B. Zahrin, N. Venkataraman, and G. S. Hiong, “Benchmarking a time-series foundation model (timegpt) for real-world forecasting applications,”Ma- chine Learning with Applications, p. 100801, 2025

2025

[5] [5]

Mira: Medical time series foundation model for real-world health data,

H. Li, B. Deng, C. Xu, Z. Feng, V . Schlegel, Y .-H. Huang, Y . Sun, J. Sun, K. Yang, Y . Yuet al., “Mira: Medical time series foundation model for real-world health data,”Advances in Neural Information Processing Systems, vol. 38, pp. 98 657– 98 685, 2026

2026

[6] [6]

How effective are large time series models in hydrology? a study on water level forecasting in everglades,

R. Rangaraj, J. Shi, A. Shirali, R. Paudel, Y . Wu, and G. Narasimhan, “How effective are large time series models in hydrology? a study on water level forecasting in everglades,” arXiv preprint arXiv:2505.01415, 2025

work page arXiv 2025

[7] [7]

From detection to forecasting: utilizing time-series foundation models to anticipate defects in metal additive manufacturing,

J. Zhang, F. Farbiz, M. Jafary-Zadeh, and S. L. Sing, “From detection to forecasting: utilizing time-series foundation models to anticipate defects in metal additive manufacturing,”Journal of Manufacturing Processes, vol. 150, pp. 1040–1052, 2025

2025

[8] [8]

Ultra low power mox sensor reading for natural gas wireless monitoring,

M. Rossi and D. Brunelli, “Ultra low power mox sensor reading for natural gas wireless monitoring,”IEEE Sensors Journal, vol. 14, no. 10, pp. 3433–3441, 2014

2014

[9] [9]

Mox-nw electronic nose for detection of food microbial contamination,

G. Sberveglieri, G. Zambotti, M. Falasconi, E. Gobbi, and V . Sberveglieri, “Mox-nw electronic nose for detection of food microbial contamination,” inSENSORS, 2014 IEEE. IEEE, 2014, pp. 1376–1379

2014

[10] [10]

Calibration transfer and drift counteraction in chemical sensor arrays using direct standardization,

J. Fonollosa, L. Fernandez, A. Guti ´errez-G´alvez, R. Huerta, and S. Marco, “Calibration transfer and drift counteraction in chemical sensor arrays using direct standardization,”Sensors and Actuators B: Chemical, vol. 236, pp. 1044–1053, 2016

2016

[11] [11]

Mdfe-net: A meta-learning driven dual-branch feature extraction network for e-nose sensor drift adaptation,

Q. Yang, J. Liu, Y . Shi, Y . Wang, and H. Men, “Mdfe-net: A meta-learning driven dual-branch feature extraction network for e-nose sensor drift adaptation,”ACS sensors, vol. 11, no. 5, pp. 4057–4067, 2026

2026