Pith. sign in

REVIEW 3 major objections 6 minor 49 references

A small trained BiLSTM beats zero-shot and LoRA-tuned foundation models at forecasting extreme wildfire PM2.5 under leave-one-incident-out evaluation.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-10 14:49 UTC pith:BCY6BHVB

load-bearing objection Careful LOIO benchmark shows a trained BiLSTM beats zero-shot and LoRA TSFMs on extreme wildfire PM2.5; the ranking holds under the stated univariate design. the 3 major comments →

arxiv 2607.07951 v1 pith:BCY6BHVB submitted 2026-07-08 cs.LG physics.ao-ph

Evaluating the Generalizability of Foundation Models for Extreme Environmental Events: Case Study of California Wildfire PM2.5

Yongcan Huang , Li Jiang , Ze Yu Liu This is my paper
classification cs.LG physics.ao-ph
keywords PM2.5 forecastingwildfire smoketime series foundation modelsleave-one-incident-outexceedance predictionLoRA fine-tuningdeep learningair quality
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Wildfire smoke produces rare, hazardous PM2.5 spikes that public-health systems need to forecast ahead of time, yet most air-quality models are tested on chronological splits that can leak information from the same fire. This paper builds a 12-year California benchmark of 1,375 wildfire incidents across 79 monitors and evaluates models with a leave-one-incident-out protocol that holds out entire fires. It compares four zero-shot time-series foundation models and two LoRA-adapted versions against fully trained LSTM, BiLSTM, and Transformer baselines plus naive persistence, measuring both ordinary error and whether hazardous AQI thresholds are correctly flagged at 6-, 12-, and 24-hour horizons. The trained BiLSTM wins every metric, including the hardest Hazardous band, while foundation models only modestly beat persistence and never overtake the recurrent baselines even after fine-tuning. The result matters because it challenges the assumption that large pretrained forecasters automatically dominate extreme environmental events and supplies concrete guidance on when a compact domain-trained model remains the better operational choice.

Core claim

Under leave-one-incident-out evaluation on California wildfire PM2.5, a fully trained BiLSTM achieves the lowest MAE (5.16 µg/m³) and the highest exceedance F1 at every EPA AQI threshold, including Hazardous (>225.5 µg/m³) at 0.63, while no zero-shot or LoRA-adapted foundation model surpasses the trained recurrent baselines on any metric.

What carries the argument

Leave-one-incident-out (LOIO) cross-validation: every window from a given wildfire–site episode is held out together so that models must generalize to an entirely unseen fire, preventing the within-incident leakage that chronological splits allow.

Load-bearing premise

That a univariate 48-hour PM2.5 history alone, without weather covariates or fire-side drivers, is a fair and sufficient input for comparing foundation models against recurrent baselines on extreme-event generalization.

What would settle it

Re-run the same LOIO benchmark after adding wind, humidity, temperature inversions, burned area, or fire radiative power as covariates; if a zero-shot or LoRA-tuned foundation model then beats BiLSTM on Hazardous exceedance F1 or MAE, the central claim is overturned.

Watch this falsifier — get emailed when new claim-graph text bears on it.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The paper benchmarks time series foundation models (zero-shot TimesFM, Chronos-2, Moirai-2, Time-MoE, plus LoRA-adapted Chronos-2 and Time-MoE) against fully trained LSTM, BiLSTM, and Transformer baselines and naïve persistence for wildfire-driven PM2.5 forecasting. Using a 12-year California panel (79 sites, 1,375 incidents) and a severity-stratified leave-one-incident-out protocol, it evaluates MAE, RMSE, R², and exceedance F1 at EPA AQI thresholds over 6/12/24-hour horizons from a shared 48-hour univariate context. The central empirical claim is that BiLSTM attains the lowest MAE (5.16 µg/m³) and the highest exceedance F1 at every threshold, including Hazardous (0.63), while no zero-shot or LoRA-adapted foundation model surpasses the trained recurrent baselines on any metric; zero-shot Chronos-2 shows severe RMSE/R² tail instability that LoRA largely repairs without closing the gap.

Significance. If the ranking holds, the work is a useful corrective to the assumption that large pretrained TSFMs automatically dominate extreme environmental forecasting. Strengths include a carefully designed LOIO protocol that avoids within-incident leakage, severity-stratified folds, shared windows across horizons, residual parameterization for trained models, transparent reporting of Chronos-2 tail instability, and operationally relevant exceedance F1 at multiple AQI thresholds. The released multi-incident California benchmark and the deployment guidance (prefer compact trained recurrent models when multi-year fire history exists; use LoRA TSFMs only as a cold-start path) are concrete contributions for air-quality and environmental ML practice.

major comments (3)
  1. Section 4.1 and Table 3: the trained baselines predict residuals from the last observed standardized value (Eqs. 6–7), while zero-shot TSFMs (except Time-MoE) receive native-scale series and emit direct median quantiles. Residual parameterization is a strong inductive bias that anchors forecasts to persistence and stabilizes heavy-tailed training; the paper does not ablate a non-residual trained baseline or residual-style adaptation for TSFMs. Without that control, part of the BiLSTM advantage may be attributable to target parameterization rather than architecture or pretraining alone. A short ablation (or explicit residual-head TSFM variant) would make the hierarchy more conclusive.
  2. Section 4.6 and Table 4: LoRA is applied only to Chronos-2 and Time-MoE, with a fixed budget (500 steps, r=8 for Time-MoE, lr=1e-5). The claim that "no foundation model, zero-shot or fine-tuned, surpasses the trained recurrent baselines" therefore rests on two adapted families only. Given that TimesFM and Moirai-2 are competitive zero-shot, either adapting them under the same LOIO protocol or justifying their exclusion is needed before the adaptation conclusion is fully general.
  3. Section 5 and Discussion: all models are strictly univariate (48 h PM2.5 only). The authors correctly note this as a limitation, but the abstract and contribution list frame the result as challenging the universal dominance of larger pretrained models in environmental forecasting. That framing is stronger than the controlled univariate comparison supports. Soften the claim to the stated setting (identical univariate inputs, LOIO wildfire PM2.5) and treat multivariate/covariate-capable TSFMs as future work rather than as already refuted.
minor comments (6)
  1. Abstract vs. Table 4: Chronos-2 zero-shot RMSE is given as 23.4 µg/m³ in the abstract and 23.45 in the table; keep one consistent rounding convention.
  2. Figure 6 includes a Moderate (9.1) threshold that is not listed among the primary AQI breakpoints in the abstract or Table 4; either add it to the main table or drop it from the figure for consistency.
  3. Section 3.2: "evaluation was conducted on held-out continuous weeks containing elevated-concentration episodes" is slightly ambiguous relative to the LOIO protocol of Section 3.3; clarify that LOIO is the sole evaluation design.
  4. Table 3 lists Moirai-2 as "patch; dec." while Section 4.5 describes it as a patch-based encoder; align the architecture description.
  5. Figure 7 caption and body: incident ID is truncated differently across places; use a single full identifier for reproducibility.
  6. A few typographic issues: "naïve" vs "naïve" consistency, and the arXiv date line "July 10, 2026" looks like a placeholder.

Circularity Check

0 steps flagged

No significant circularity: the BiLSTM-over-TSFM ranking is an empirical LOIO comparison, not a derivation that reduces to its inputs by construction.

full rationale

This paper is a controlled empirical benchmark, not a first-principles derivation. Its central claim is a performance hierarchy under leave-one-incident-out evaluation (Section 3.3, Table 4, Figures 4–6): BiLSTM achieves the lowest MAE (5.16 µg/m³) and highest exceedance F1 at every EPA AQI threshold, while no zero-shot or LoRA-adapted TSFM surpasses the trained recurrent baselines. The LOIO protocol groups all windows from a wildfire–site episode into a single fold so held-out fires have no training counterpart; residual targets for trained baselines are inverted to native units for evaluation; persistence is reported as an explicit non-learning floor; and LoRA adapters are trained only on each fold’s training incidents. None of these steps defines a claimed prediction in terms of a fitted constant that is then re-presented as independent skill. There is no self-definitional equation, no fitted parameter renamed as a prediction, no load-bearing uniqueness theorem imported from the authors, and no ansatz smuggled via self-citation. The univariate 48-hour input limitation is a genuine scope choice already acknowledged in the Discussion, but it does not create circularity in the ranking as stated. Score 0 is therefore the correct, proportionate finding.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 0 invented entities

The central claim is an empirical ranking under a defined protocol; it rests on standard ML and air-quality assumptions plus a few design choices that function as free parameters. No new physical entities are postulated.

free parameters (6)
  • input context length Lin = 48 h
    Chosen by the authors; not derived from first principles and directly shapes every model’s receptive field.
  • forecast horizons {6,12,24} h and stride 6 h
    Design choices that define the supervised examples and the reported metrics.
  • backbone width d=64, two layers, dropout 0.1 for trained baselines
    Architecture hyperparameters fixed by the authors and shared across LSTM/BiLSTM/Transformer.
  • LoRA rank r=8, alpha=16, 500 steps, lr=1e-5 for adapted models
    Adaptation budget and rank chosen by hand; different values could alter the fine-tuning gap.
  • 100 km station-to-incident distance cutoff
    Ad-hoc geographic filter that determines which incidents enter the panel.
  • severity-quintile stratification for LOIO folds
    Fold construction rule that balances the long-tailed severity distribution; alternative stratifications could change fold difficulty.
axioms (5)
  • domain assumption Leave-one-incident-out grouping prevents temporal leakage and is the correct operational generalization test for wildfire events.
    Stated in Section 3.3 and used as the sole evaluation protocol; if chronological splits were equally valid the ranking could change.
  • domain assumption Univariate PM2.5 history is a fair common input for comparing foundation models and recurrent baselines.
    Section 4.1; multivariate covariates are deferred to future work, yet the claim is framed as a general TSFM-vs-trained comparison.
  • domain assumption Predictive median of quantile heads is the appropriate point forecast for MAE/RMSE/F1 comparison.
    Section 4.5; alternative quantiles or scoring rules could reorder models on exceedance metrics.
  • ad hoc to paper Per-site z-score standardization (trained models) versus native-scale instance normalization (most TSFMs) does not unfairly advantage either class.
    Table 3 and Section 4.1; the paper asserts architectural fairness but the scaling difference remains a design choice.
  • standard math Standard deep-learning training (MSE on residuals, early stopping) and public pretrained TSFM interfaces behave as documented.
    Background ML practice assumed throughout Sections 4.4–4.6.

pith-pipeline@v1.1.0-grok45 · 25023 in / 3118 out tokens · 47462 ms · 2026-07-10T14:49:34.419473+00:00 · methodology

0 comments
read the original abstract

Wildfire smoke events produce extreme PM$_{2.5}$ concentrations that pose severe public health risks, yet forecasting rare, hazardous-level spikes remains a fundamental challenge. Time series foundation models (TSFMs), pretrained models offering zero-shot inference and efficient adaptation, perform strongly on general benchmarks, but their behavior under extreme out-of-distribution conditions is poorly understood. We present the first systematic benchmark comparing six TSFM configurations (zero-shot TimesFM, Chronos-2, Moirai-2, and Time-MoE, plus LoRA fine-tuned Chronos-2 and Time-MoE) against fully-trained baselines (LSTM, BiLSTM, Transformer) and naive persistence on a 12-year (2013--2025) hourly PM$_{2.5}$ dataset covering 1,375 wildfire incidents across 79 California monitoring sites. A leave-one-incident-out (LOIO) protocol evaluates generalization to unseen fires, using MAE, RMSE, and exceedance F1 at EPA AQI thresholds across 6-, 12-, and 24-hour horizons. Results reveal a consistent hierarchy. The BiLSTM achieves the lowest MAE ($5.16\,\mu g/m^3$) and the highest exceedance F1 at every threshold, including the Hazardous band ($>225.5\,\mu g/m^3$), reaching 0.63 versus at most 0.54 for any foundation model. Zero-shot TSFMs improve on persistence only modestly, and zero-shot Chronos-2 exhibits severe RMSE tail instability ($23.4\,\mu g/m^3$, negative $R^2$) from sporadic large errors. LoRA fine-tuning substantially improves both adapted families and largely repairs this instability, yet no foundation model surpasses the trained recurrent baselines on any metric. These findings challenge the assumption that larger pretrained models universally dominate environmental forecasting and provide actionable deployment guidance for wildfire air quality prediction.

Figures

Figures reproduced from arXiv: 2607.07951 by Li Jiang, Yongcan Huang, Ze Yu Liu.

Figure 1
Figure 1. Figure 1: Spatial distribution of the 79 EPA-certified PM [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Distribution of hourly PM [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Comparison of data-splitting strategies. (a) A naive chronological split places early [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Overall mean absolute error by model, averaged over five folds and the 6-, 12-, and [PITH_FULL_IMAGE:figures/full_fig_p015_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: (a) MAE and (b) RMSE as a function of forecast horizon, averaged over five folds. [PITH_FULL_IMAGE:figures/full_fig_p016_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Exceedance-detection F1 across the 24-hour PM [PITH_FULL_IMAGE:figures/full_fig_p017_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: PM2.5 forecasts on a held-out severe incident (incident 1d1bb813, site 3132) at 6-, 12-, and 24-hour lead times (top to bottom), issued at every hour and split into the main surge (10–120 h) and decay phase (120–200 h). Shaded bands mark the Unhealthy, Very Unhealthy, and Hazardous AQI categories; the near-baseline tail beyond 200 h is omitted for clarity. Peak under-prediction and timing misalignment grow… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

49 extracted references · 49 canonical work pages · 2 internal anchors

  1. [1]

    Time-series large language models: A systematic review of state-of-the-art.IEEE Access, 13:30235–30261, 2025

    Shamsu Abdullahi, Kamaluddeen Usman Danyaro, Abubakar Zakari, Izzatdin Abdul Aziz, Noor Amila Wan Abdullah Zawawi, and Shamsuddeen Adamu. Time-series large language models: A systematic review of state-of-the-art.IEEE Access, 13:30235–30261, 2025. doi: 10.1109/ACCESS.2025.3535782

    work page doi:10.1109/access.2025.3535782 2025

  2. [2]

    Ensemble-based deep learning for estimating PM2.5 over California with multisource big data including wildfire smoke.Environment International, 155:106721, 2021

    Rosana Aguilera, Thomas Corringham, Alexander Gershunov, and Tarik Benmarhnia. Ensemble-based deep learning for estimating PM2.5 over California with multisource big data including wildfire smoke.Environment International, 155:106721, 2021. doi: 10.1016/j.envint.2021.106721

    work page doi:10.1016/j.envint.2021.106721 2021

  3. [3]

    Using VIIRS fire radiative power data to simulate biomass burning emissions, plume rise and smoke transport in a real-time air quality modeling system

    Ravan Ahmadov, Georg Grell, Eric James, Ivan Csiszar, Marina Tsidulko, Brad Pierce, Stuart McKeen, Stan Benjamin, Curtis Alexander, Gabriel Pereira, et al. Using VIIRS fire radiative power data to simulate biomass burning emissions, plume rise and smoke transport in a real-time air quality modeling system. In2017 IEEE International Geoscience and Remote S...

    work page 2017

  4. [4]

    Chronos: Learning the Language of Time Series

    Abdul Fatir Ansari, Lorenzo Stella, Caner Turkmen, Xiyuan Zhang, Pedro Mercado, Huibin Shen, Oleksandr Shchur, Syama Sundar Rangapuram, Sebastian Pineda Arango, Shubham Kapoor, Jasper Zschiegner, Danielle C. Maddix, Michael W. Mahoney, Kari Torkkola, Andrew Gordon Wilson, Michael Bohlke-Schneider, and Yuyang Wang. Chronos: Learning the language of time se...

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  5. [5]

    Chronos-2: From Univariate to Universal Forecasting

    Abdul Fatir Ansari, Oleksandr Shchur, Jaris K¨ uken, Andreas Auer, Boran Han, Pedro Mer- cado, Syama Sundar Rangapuram, Huibin Shen, Lorenzo Stella, Xiyuan Zhang, Mononito Goswami, Shubham Kapoor, et al. Chronos-2: From univariate to universal forecasting. arXiv preprint arXiv:2510.15821, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  6. [6]

    Bhowmik, Youn Soo Jung, Juan A

    Rohan T. Bhowmik, Youn Soo Jung, Juan A. Aguilera, Mary Prunicki, and Kari Nadeau. A multi-modal wildfire prediction and early-warning system based on a novel machine learning framework.Journal of Environmental Management, 341:117908, 2023

    work page 2023

  7. [7]

    Cross-regional deep learning for air quality forecasting: A comparative study of co, no2, o3, pm2

    Adam Booth, Philip James, Stephen McGough, and Ellis Solaiman. Cross-regional deep learning for air quality forecasting: A comparative study of co, no2, o3, pm2. 5, and pm10. Forecasting, 7(4):66, 2025

    work page 2025

  8. [8]

    The changing risk and burden of wildfire in the United States.Proceedings of the National Academy of Sciences, 118(2), 2021

    Marshall Burke, Anne Driscoll, Sam Heft-Neal, et al. The changing risk and burden of wildfire in the United States.Proceedings of the National Academy of Sciences, 118(2), 2021

    work page 2021

  9. [9]

    Daewon Byun and Kenneth L. Schere. Review of the governing equations, computational algorithms, and other components of the models-3 community multiscale air quality (CMAQ) modeling system.Applied Mechanics Reviews, 59(2):51–77, 2006

    work page 2006

  10. [10]

    This time is different: An observability perspective on time series foundation models.Advances in neural information processing systems, 38:50907–50951, 2026

    Ben Cohen, Emaad Khwaja, Youssef Doubli, Salahidine Lemaachi, Chris Lettieri, Charles Masson, Hugo Miccinilli, Elise Ram´ e, Qiqi Ren, Afshin Rostamizadeh, et al. This time is different: An observability perspective on time series foundation models.Advances in neural information processing systems, 38:50907–50951, 2026. 20

    work page 2026

  11. [11]

    A decoder-only foundation model for time-series forecasting

    Abhimanyu Das, Weihao Kong, Rajat Sen, and Yichen Zhou. A decoder-only foundation model for time-series forecasting. InProceedings of the 41st International Conference on Machine Learning (ICML), volume 235 ofProceedings of Machine Learning Research, pages 10148–10167. PMLR, 2024

    work page 2024

  12. [12]

    Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity.Journal of Machine Learning Research, 23(120):1–39, 2022

    William Fedus, Barret Zoph, and Noam Shazeer. Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity.Journal of Machine Learning Research, 23(120):1–39, 2022

    work page 2022

  13. [13]

    Freeman, Graham Taylor, Bahram Gharabaghi, and Jesse Th´ e

    Brian S. Freeman, Graham Taylor, Bahram Gharabaghi, and Jesse Th´ e. Forecasting air quality time series using deep learning.Journal of the Air & Waste Management Association, 68(8):866–886, 2018

    work page 2018

  14. [14]

    MOMENT: A family of open time-series foundation models

    Mononito Goswami, Konrad Szafer, Arjun Choudhry, Yifu Cai, Shuo Li, and Artur Dubrawski. MOMENT: A family of open time-series foundation models. InProceed- ings of the 41st International Conference on Machine Learning (ICML), 2024

    work page 2024

  15. [15]

    Grell, Steven E

    Georg A. Grell, Steven E. Peckham, Rainer Schmitz, et al. Fully coupled online chemistry within the WRF model.Atmospheric Environment, 39(37):6957–6975, 2005

    work page 2005

  16. [16]

    Long short-term memory.Neural Computation, 9(8):1735–1780, 1997

    Sepp Hochreiter and J¨ urgen Schmidhuber. Long short-term memory.Neural Computation, 9(8):1735–1780, 1997

    work page 1997

  17. [17]

    Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen

    Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations (ICLR), 2022

    work page 2022

  18. [18]

    Jaffe, Susan M

    Daniel A. Jaffe, Susan M. O’Neill, Narasimhan K. Larkin, Amara L. Holder, David L. Peterson, Jessica E. Halofsky, and Ana G. Rappold. Wildfire and prescribed burning impacts on air quality in the United States.Journal of the Air & Waste Management Association, 70(6):583–615, 2020

    work page 2020

  19. [19]

    A. K. Kochanski, M. A. Jenkins, J. Mandel, J. D. Beezley, and S. K. Krueger. Toward an integrated system for fire, smoke and air quality simulations.Geoscientific Model Development, 6(4):1071–1084, 2013. doi: 10.5194/gmd-6-1071-2013

    work page doi:10.5194/gmd-6-1071-2013 2013

  20. [20]

    Krishan Kumar and B. P. Pande. Air pollution prediction with machine learning: A case study of Indian cities.International Journal of Environmental Science and Technology, 20: 5333–5348, 2023. doi: 10.1007/s13762-022-04241-5

    work page doi:10.1007/s13762-022-04241-5 2023

  21. [21]

    Learning skillful medium-range global weather forecasting.Science, 382(6677):1416–1421, 2023

    Remi Lam, Alvaro Sanchez-Gonzalez, Matthew Willson, Peter Wirnsberger, Meire Fortunato, Ferran Alet, Suman Ravuri, Timo Ewalds, Zach Eaton-Rosen, Weihua Hu, et al. Learning skillful medium-range global weather forecasting.Science, 382(6677):1473–1479, 2023. doi: 10.1126/science.adi2336

    work page doi:10.1126/science.adi2336 2023

  22. [22]

    Multi-scale spectral recurrent network based on random fourier features for wind speed forecasting

    Eder Arley Leon-Gomez, V´ ıctor Elvira, Jorge Iv´ an Montes-Monsalve, Andr´ es Marino ´Alvarez-Meza, Alvaro Orozco-Gutierrez, and German Castellanos-Dominguez. Multi-scale spectral recurrent network based on random fourier features for wind speed forecasting. Technologies, 14(4):238, 2026

    work page 2026

  23. [23]

    Long short-term memory neural network for air pollutant concentration predictions: Method development and evaluation.Environmental Pollution, 231:997–1004, 2017

    Xiang Li, Ling Peng, Xiaojing Yao, et al. Long short-term memory neural network for air pollutant concentration predictions: Method development and evaluation.Environmental Pollution, 231:997–1004, 2017. 21

    work page 2017

  24. [24]

    Y. Li, D. Q. Tong, F. Ngan, M. D. Cohen, A. F. Stein, S. Kondragunta, X. Zhang, C. Ichoku, E. J. Hyer, and R. A. Kahn. Ensemble PM2.5 forecasting during the 2018 Camp Fire event using the HYSPLIT transport and dispersion model.Journal of Geophysical Research: Atmospheres, 125(15):e2020JD032768, 2020

    work page 2018

  25. [25]

    Foundation models for time series analysis: A tutorial and survey

    Yuxuan Liang, Haomin Wen, Yuqi Nie, Yushan Jiang, Ming Jin, Dongjin Song, Shirui Pan, and Qingsong Wen. Foundation models for time series analysis: A tutorial and survey. InProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD), pages 6555–6565, 2024. doi: 10.1145/3637528.3671451

    work page doi:10.1145/3637528.3671451 2024

  26. [26]

    Lamb, and Pierre Gentine

    Kyleen Liao, Jatan Buch, Kara D. Lamb, and Pierre Gentine. Simulating the air quality impact of prescribed fires using graph neural network-based PM2.5 forecasts.Environmental Data Science, 4:e10, 2025. doi: 10.1017/eds.2025.4

    work page doi:10.1017/eds.2025.4 2025

  27. [27]

    Moirai 2.0: When less is more for time series forecasting.arXiv preprint arXiv:2511.11698, 2025

    Chenghao Liu, Taha Aksu, Juncheng Liu, Xu Liu, Hanshu Yan, Quang Pham, Silvio Savarese, Doyen Sahoo, Caiming Xiong, and Junnan Li. Moirai 2.0: When less is more for time series forecasting.arXiv preprint arXiv:2511.11698, 2025

    work page arXiv 2025

  28. [28]

    Timer: Generative pre-trained transformers are large time series models

    Yong Liu, Haoran Zhang, Chenyu Li, Xiangdong Huang, Jianmin Wang, and Mingsheng Long. Timer: Generative pre-trained transformers are large time series models. InProceed- ings of the 41st International Conference on Machine Learning (ICML), 2024

    work page 2024

  29. [29]

    Sundial: A family of highly capable time series foundation models, 2025

    Yong Liu, Guo Qin, Zhiyuan Shi, Zhi Chen, Caiyin Yang, Xiangdong Huang, Jianmin Wang, and Mingsheng Long. Sundial: A family of highly capable time series foundation models, 2025

    work page 2025

  30. [30]

    Z. Liu, Z. Fang, and Y. Hu. A deep learning-based hybrid method for PM2.5 prediction in cen- tral and western China.Scientific Reports, 15:10080, 2025. doi: 10.1038/s41598-025-95460-6

    work page doi:10.1038/s41598-025-95460-6 2025

  31. [31]

    From pre-training to post-training: A survey on time series foundation models.TechRxiv,

    Zhen Liu, Boyuan Li, Hao Huang, Yanru Sun, Yucheng Wang, Min Wu, and Qianli Ma. From pre-training to post-training: A survey on time series foundation models.TechRxiv,

    work page

  32. [32]

    doi: 10.36227/techrxiv.176978429.90235801

    work page doi:10.36227/techrxiv.176978429.90235801

  33. [33]

    Qianli Ma, Zhen Liu, Zhenjing Zheng, Ziyang Huang, Siying Zhu, Zhongzhong Yu, and James T. Kwok. A survey on time-series pre-trained models.IEEE Transactions on Knowledge and Data Engineering, 36(12):7536–7555, 2024. doi: 10.1109/TKDE.2024. 3475809

    work page doi:10.1109/tkde.2024 2024

  34. [34]

    Lag-Llama: Towards foundation models for probabilistic time series forecasting, 2023

    Kashif Rasul, Arjun Ashok, Andrew Robert Williams, Hena Ghonia, Rishika Bhagwatkar, Arian Khorasani, Mohammad Javad Darvishi Bayazi, George Adamopoulos, Roland Riachi, Nadhir Hassen, Marin Biloˇ s, Sahil Garg, Anderson Schneider, Nicolas Chapados, Alexandre Drouin, Valentina Zantedeschi, Yuriy Nevmyvaka, and Irina Rish. Lag-Llama: Towards foundation model...

    work page 2023

  35. [35]

    Reid, Michael Brauer, Fay H

    Colleen E. Reid, Michael Brauer, Fay H. Johnston, et al. Critical review of health impacts of wildfire smoke exposure.Environmental Health Perspectives, 124(9):1334–1343, 2016

    work page 2016

  36. [36]

    Mar´ ıa Inmaculada Rodr´ ıguez-Garc´ ıa, Mar´ ıa Gema Carrasco-Garc´ ıa, Paloma Roc´ ıo Cu- billas Fern´ andez, Maria da Concei¸ c˜ ao Rodrigues Ribeiro, Pedro J. S. Cardoso, and Ig- nacio J. Turias. Air pollution forecasting using autoencoders: A classification-based prediction of NO2, PM10, and SO2 concentrations.Nitrogen, 6(4):101, 2025. doi: 10.3390/n...

    work page doi:10.3390/nitrogen6040101 2025

  37. [37]

    M. K. Saravana, M. S. Roopa, J. S. Arunalatha, and K. R. Venugopal. Transform- ers for multivariate time series forecasting: Comprehensive analysis, challenges, re- search opportunities and future prospects.IEEE Access, 14:11424–11457, 2026. doi: 10.1109/ACCESS.2026.3654408. 22

    work page doi:10.1109/access.2026.3654408 2026

  38. [38]

    Time-MoE: Billion-scale time series foundation models with mixture of experts

    Xiaoming Shi, Shiyu Wang, Yuqi Nie, Dianqi Li, Zhou Ye, Qingsong Wen, and Ming Jin. Time-MoE: Billion-scale time series foundation models with mixture of experts. InThe Thirteenth International Conference on Learning Representations (ICLR), 2025

    work page 2025

  39. [39]

    A. F. Stein, R. R. Draxler, G. D. Rolph, B. J. B. Stunder, M. D. Cohen, and F. Ngan. NOAA’s HYSPLIT atmospheric transport and dispersion modeling system.Bulletin of the American Meteorological Society, 96(12):2059–2077, 2015. doi: 10.1175/BAMS-D-14-00110.1

    work page doi:10.1175/bams-d-14-00110.1 2059

  40. [40]

    Gomez, Lukasz Kaiser, and Illia Polosukhin

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. InAdvances in Neural Information Processing Systems (NeurIPS), volume 30, 2017

    work page 2017

  41. [41]

    ChatTime: A unified multimodal time series foundation model bridging numerical and textual data

    Chengsen Wang, Qi Qi, Jingyu Wang, Haifeng Sun, Zirui Zhuang, Jinming Wu, Lei Zhang, and Jianxin Liao. ChatTime: A unified multimodal time series foundation model bridging numerical and textual data. InProceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 12694–12702, 2025

    work page 2025

  42. [42]

    Unified training of universal time series forecasting transformers

    Gerald Woo, Chenghao Liu, Akshat Kumar, Caiming Xiong, Silvio Savarese, and Doyen Sahoo. Unified training of universal time series forecasting transformers. InProceedings of the 41st International Conference on Machine Learning (ICML), volume 235 ofProceedings of Machine Learning Research. PMLR, 2024. Oral presentation

    work page 2024

  43. [43]

    World Health Organization, Geneva, 2021

    World Health Organization.WHO Global Air Quality Guidelines: Particulate Matter (PM2.5 and PM10), Ozone, Nitrogen Dioxide, Sulfur Dioxide and Carbon Monoxide. World Health Organization, Geneva, 2021

    work page 2021

  44. [44]

    Spatial deep learning for PM2.5 estimation during extreme pollution events: Quantifying uncertainty and training with sparse datasets.Environmental Research: Atmospheres, 2026

    Yue Wu et al. Spatial deep learning for PM2.5 estimation during extreme pollution events: Quantifying uncertainty and training with sparse datasets.Environmental Research: Atmospheres, 2026. doi: 10.1088/3049-4753/ae3c12

    work page doi:10.1088/3049-4753/ae3c12 2026

  45. [45]

    X. Ye, P. Arab, R. Ahmadov, E. James, G. A. Grell, B. Pierce, et al. Evaluation and intercomparison of wildfire smoke forecasts from multiple modeling systems for the 2019 Williams Flats fire.Atmospheric Chemistry and Physics, 21(18):14427–14469, 2021. doi: 10.5194/acp-21-14427-2021

    work page doi:10.5194/acp-21-14427-2021 2019

  46. [46]

    Predicting hourly PM2.5 concentrations in wildfire-prone areas using a spatiotemporal transformer model.Science of The Total Environment, 860:160446, 2023

    Manzhu Yu, Arif Masrur, and Christopher Blaszczak-Boxe. Predicting hourly PM2.5 concentrations in wildfire-prone areas using a spatiotemporal transformer model.Science of The Total Environment, 860:160446, 2023

    work page 2023

  47. [47]

    Evaluating deep learning time series models for PM2.5 forecasting across diverse horizons.iScience, 29(2): 114770, 2026

    Ling Zeng, Runan Dong, Meng Yuan, Linhai Jing, and Shoutao Jiao. Evaluating deep learning time series models for PM2.5 forecasting across diverse horizons.iScience, 29(2): 114770, 2026. doi: 10.1016/j.isci.2026.114770

    work page doi:10.1016/j.isci.2026.114770 2026

  48. [48]

    A spatiotemporal multimodal framework for air pollution prediction based on bayesian optimization: Evidence from Sichuan, China

    Fengfan Zhang, Jiabei Hu, and Ming Zeng. A spatiotemporal multimodal framework for air pollution prediction based on bayesian optimization: Evidence from Sichuan, China. Atmosphere, 16(8):958, 2025. doi: 10.3390/atmos16080958

    work page doi:10.3390/atmos16080958 2025

  49. [49]

    Real-time air quality forecasting, part I: History, techniques, and current status.Atmospheric Environment, 60:632–655, 2012

    Yang Zhang, Marc Bocquet, Vivien Mallet, et al. Real-time air quality forecasting, part I: History, techniques, and current status.Atmospheric Environment, 60:632–655, 2012. 23

    work page 2012