PIER: Physics-Informed Environmental Retrieval for Time-Series Modeling

Source: arXiv:2607.20230 · Published 2026-07-22 · By Shiyuan Luo, Runlong Yu, Chonghao Qiu, Yue Qin, Rahul Ghosh, Robert Ladwig et al.

TL;DR

This paper addresses the challenge of modeling heterogeneous environmental time series data, such as water temperature and dissolved oxygen (DO) concentrations in lakes, where data are sparse and physical dynamics vary across systems. Traditional embedding-based retrieval approaches can transfer knowledge across systems but risk retrieving physically inconsistent scenarios due to similar embeddings arising from different underlying mechanisms. The authors propose Physics-Informed Environmental Retrieval (PIER), a model-agnostic framework which augments standard embedding-based retrieval with a physics-aware retrieval stream. This physics stream leverages local verifiers trained on physics-derived flux features to score candidate scenarios by how consistently their physical flux responses match the target system’s dynamics. An adaptive weighting mechanism learns per-scenario weights to dynamically balance the embedding similarity and physics-informed scores based on reliability diagnostics. Evaluations on a comprehensive dataset of 356 lakes from the Midwestern U.S. over 41 years demonstrate that PIER consistently improves water temperature and DO prediction accuracy compared to baselines, including physics-only models and advanced deep learning architectures. The framework also generalizes across diverse predictive backbones.

Key findings

• PIER reduces RMSE in dissolved oxygen concentration prediction by 4.73% to 12.4% compared to the best baseline across six seasonal settings (e.g., 1.8888 vs 1.9841 for epilimnion DO).
• PIER lowers water temperature RMSE by 11.9% to 15.4% over baselines (e.g., 1.2486 vs 1.5043 for summer epilimnion temperature).
• Advanced architectures like iTransformer and TSMixer perform worse than a simpler LSTM, highlighting data sparsity challenges.
• PIER’s physics-aware retrieval stream promotes scenarios with physically consistent flux-response dynamics, improving accuracy especially for atypical conditions that embedding similarity retrieval misrepresents.
• Adaptive weight adjustment per scenario (based on diagnostics such as local verifier quality) outperforms fixed or single-stream retrieval strategies.
• Ablation study shows combined retrieval with adaptive weighting (PIER) outperforms methods using only embedding similarity or physics-aware scoring alone.
• PIER effectively integrates physics-based process model-derived flux features with data-driven embeddings to balance transfer learning and system specificity.
• The framework is backbone-agnostic and consistently improves performance across models including LSTM, Informer, and TimesNet.

Methodology — deep read

1. Threat Model & Assumptions: The adversary is not explicitly defined as this work focuses on environmental modeling. Assumptions include availability of partial environmental observations and process-based model simulations providing physics-derived flux features. The main challenge is dealing with sparse and heterogeneous time series where direct observation is insufficient for accurate modeling.

2. Data: The authors use a large ecological dataset comprising ~1.75 million daily records from 356 lakes across the Midwestern US from 1979-2019. The data include 47 environmental features (e.g., morphometry, meteorology, trophic state). Observations of water temperature and dissolved oxygen are recorded on subsets of days with varying coverage. Data split into training (up to 2011), validation (2012-2015), and testing (2016-2019). Physics-derived flux variables are generated from process-based models (General Lake Model for heat flux; specific biogeochemical models for DO).

3. Architecture / Algorithm: PIER is a two-stream retrieval-augmented modeling framework. A global predictive model (typically an LSTM) is trained across all scenarios, mapping each scenario to a latent embedding. Retrieval is augmented by a physics-aware stream: for each system, a local verifier is trained using only physics-derived flux features to model flux-response dynamics. Candidates from embedding similarity and physics streams are scored separately; physics scores come from applying the target system’s local verifier to candidate flux data and measuring predictive error converted to a similarity score via an exponential function. A weight parameter γ per target scenario governs how combined retrieval scores balance these two.

4. Training Regime: The global model is pretrained on all scenarios using environmental inputs and simulation labels from the physics models. The local verifiers are independently trained for each system on their flux features and labeled observations. The weight adjustment mechanism learns a small MLP that, given scenario diagnostic features (local verifier quality, candidate pass rate, stream agreement), predicts γ to optimize downstream prediction accuracy. This is done via a bilevel optimization approach: inner loop adapts a scenario-specific predictor fine-tuned on retrieved data weighted by combined scores; outer loop identifies γ values minimizing target scenario prediction error across a discrete candidate set for training, enabling training of the MLP gate.

5. Evaluation Protocol: Performance is measured by RMSE on predictions of water temperature and dissolved oxygen concentration across multiple seasonal partitions (e.g., summer epilimnion, hypolimnion, fall to spring). Baseline comparisons include physics-based model alone, LSTM, and several recent time series forecasting architectures. Ablations test the utility of each component: simple retrieval, physics-aware retrieval alone, fixed weighting, representation similarity alone, and full PIER. Case studies qualitatively examine promoted and demoted retrieval candidates.

6. Reproducibility: The paper notes use of publicly documented datasets and process-based models but does not specify public code availability. Hyperparameters and evaluation splits are described in detail. The local verifier training and physics derived flux computations depend on nontrivial environmental domain knowledge and simulation outputs, which may challenge direct replication. Overall, the methodology is clearly described but full reproducibility would require access to processed flux features and local verifier training pipelines.

End-to-end example: For a target lake-year scenario, the global model produces an embedding. Candidate lake-year scenarios are retrieved by embedding similarity and independently scored by physical-consistency using the target lake’s local verifier applied to candidate flux data. Candidates from both streams are merged. The learned weight adjustment MLP predicts a scenario-specific γ to combine scores. The top-K combined scored candidates form the retrieval set. A scenario-specific predictor is initialized from the global model and fine-tuned on weighted data from these retrieved scenarios, producing the final time-series predictions.

Technical innovations

• Introduction of a physics-aware retrieval stream that scores candidate scenarios based on consistency of physics-derived flux-response dynamics using local verifier models, complementing embedding-based retrieval.
• Design of a per-scenario adaptive weighting mechanism that uses diagnostic features to balance embedding similarity and physics-aware retrieval streams dynamically for optimal prediction performance.
• Formulation of a bilevel optimization framework wherein the outer loop selects retrieval stream weights to minimize target prediction error after inner loop adaptation on retrieved data.
• System-agnostic local verifier bank trained on physics model-derived flux features, enabling physics-informed filtering of retrieved scenarios for heterogeneous environmental systems.

Datasets

• Midwestern US lakes dataset — ~1.75 million daily records from 356 lakes spanning 1979–2019 — compiled from USGS, Water Quality Portal, HydroLAKES, National Land Cover Database, and published ecological sources

Baselines vs proposed

• Physics-based models (General Lake Model for temperature, process-based DO model): RMSE (DO epi.): 2.1018 vs PIER: 1.8888
• LSTM: RMSE (DO epi.): 1.9841 vs PIER: 1.8888
• Informer: RMSE (DO epi.): 2.0903 vs PIER: 1.8888
• iTransformer: RMSE (DO epi.): 2.5965 vs PIER: 1.8888
• TSMixer: RMSE (DO epi.): 2.2486 vs PIER: 1.8888
• TimesNet: RMSE (DO epi.): 2.0879 vs PIER: 1.8888
• For water temperature summer epilimnion: LSTM RMSE 1.5043 vs PIER 1.2486

Figures from the paper

Figures are reproduced from the source paper for academic discussion. Original copyright: the paper authors. See arXiv:2607.20230.

Fig 1: Overview of PIER.

Fig 2: Map of 356 tested lakes.

Fig 3: Comparison under three training settings.

Fig 4 (page 4).

Fig 5 (page 4).

Fig 4: Time-series predictions for two case study

Fig 5: Observations of promoted and demoted can-

Limitations

• PIER relies on availability and quality of physics-derived flux features from process-based models, which may not be available or accurate in all domains.
• Local verifiers can be unreliable for systems with very sparse observations, leading to deactivation of the physics-aware stream in those cases.
• The approach currently evaluates on lake ecosystems; generalization to other environmental systems is not empirically demonstrated.
• Adaptive weight learning requires discretizing γ and performing multiple inner-loop adaptations per scenario during training, which may be computationally expensive.
• While improved over baselines, the method still depends on fine-tuning per scenario, potentially limiting scalability.
• The framework’s performance gains may diminish when base predictive backbones struggle with underlying dynamics, as in some cases with iTransformer.

Open questions / follow-ons

• How can physics-aware retrieval be extended to domains lacking detailed or reliable process-based models for flux feature derivation?
• Can the adaptive weighting mechanism be further optimized or learned end-to-end with differentiable inner-loop adaptation to improve efficiency?
• How does PIER perform under real distribution shifts caused by climate change or anthropogenic impacts altering physical dynamics?
• What are the trade-offs in retrieval candidate set size and local verifier complexity for environments with limited computational resources?

Why it matters for bot defense

PIER’s framework highlights the importance of combining domain-knowledge-informed consistency checks with learned embeddings when performing retrieval-augmented modeling under sparse, heterogeneous data. For bot-defense or CAPTCHA-related time series modeling, PIER suggests that leveraging domain-specific dynamics or invariant mechanistic constraints can improve retrieval quality and downstream predictions. Its adaptive weighting of retrieval streams underscores that no single similarity measure suffices across all scenarios and that diagnostic information should guide retrieval fusion. This approach informs CAPTCHA and bot-defense practitioners to incorporate additional consistency validations beyond embedding similarity to avoid semantically misleading retrievals, especially in security contexts where adversarial or atypical data can degrade naive similarity-based retrieval methods. However, the domain-specific physics models here do not directly translate to bot-defense, so analogous invariant or mechanistically grounded features would be needed for practical adaptation.

Cite

@article{arxiv2607_20230,
  title={ PIER: Physics-Informed Environmental Retrieval for Time-Series Modeling },
  author={ Shiyuan Luo and Runlong Yu and Chonghao Qiu and Yue Qin and Rahul Ghosh and Robert Ladwig and Paul C. Hanson and Yiqun Xie and Xiaowei Jia },
  journal={arXiv preprint arXiv:2607.20230},
  year={ 2026 },
  url={https://arxiv.org/abs/2607.20230}
}

Read the full paper

• arXiv:2607.20230 (abstract)
• arXiv:2607.20230 (PDF)