# Adaptive Frequency Filtering for Groundwater Signal Separation: A Deep Learning Approach with Spectral Overlap Robustness

## Metadata

- **SPORE ID**: SPR-2026-71A1
- **Domaines**: Computer Vision x Hydrology
- **Date de generation**: 2026-08-08
- **Panel consensus score**: 6.2/10
- **Novelty score**: 0.7/1.0
- **Panel verdict**: publish_brief

## Abstract

If a learnable adaptive frequency filtering layer (AFFL) is applied to groundwater level time series, then it will achieve superior separation of anthropogenic pumping and climatic signals compared to fixed filters and standard deep learning baselines, because the per-sample spectral weighting optimizes the frequency decomposition for the specific downstream task, even under conditions of spectral overlap.

The proposed mechanism involves 5 causal steps: (1) Input groundwater level time series (1D) is transformed into the frequency domai -> (2) A learnable filter, parameterized by a small convolutional network, generates a  -> (3) The weighted frequency representation is inverse-transformed to the time domain, -> (4) The filtered signal is fed into a downstream task head (e.g., a regression layer.

Literature grounding on 7 verified references yields a novelty score of 0.7 (novel). A 3-phase experimental protocol (budget: €25k-120k, timeline: 10-16 months) is proposed, starting with in silico validation. A panel of 5 expert reviewers reached a consensus score of 6.2/10.

## 1. Hypothese et mecanisme propose

### 1.1 Formulation formelle

If a learnable adaptive frequency filtering layer (AFFL) is applied to groundwater level time series, then it will achieve superior separation of anthropogenic pumping and climatic signals compared to fixed filters and standard deep learning baselines, because the per-sample spectral weighting optimizes the frequency decomposition for the specific downstream task, even under conditions of spectral overlap.

### 1.2 Variables

**Variables independantes :**

| Variable | Type | Plage | Unite |
|----------|------|-------|-------|
| Filtering method | categorical | AFFL, Wavelet, EMD, LSTM, Linear Regression | N/A |
| Spectral overlap ratio (SOR) | continuous | 0.0 - 1.0 | dimensionless |
| Signal-to-noise ratio (SNR) | continuous | 5 - 30 | dB |

**Variables dependantes :**

| Variable | Type | Direction attendue | Unite |
|----------|------|-------------------|-------|
| Signal separation accuracy (SSA) | continuous | increase | RMSE (m) |
| Trend estimation error (TEE) | continuous | decrease | m/year |
| Recharge estimation error (REE) | continuous | decrease | mm/year |

### 1.3 Chaine causale

1. Step 1: Input groundwater level time series (1D) is transformed into the frequency domain via a Short-Time Fourier Transform (STFT) with a window size of 365 days and a hop length of 30 days.
1. Step 2: A learnable filter, parameterized by a small convolutional network, generates a per-sample, per-frequency weighting mask in the frequency domain, conditioned on the input signal's spectral envelope.
1. Step 3: The weighted frequency representation is inverse-transformed to the time domain, producing a filtered signal that emphasizes task-relevant components (e.g., pumping signal) and attenuates others (e.g., climatic noise).
1. Step 4: The filtered signal is fed into a downstream task head (e.g., a regression layer for trend estimation or a classification layer for pumping detection), and the entire pipeline is trained end-to-end via backpropagation.
1. Step 5: The loss function includes a reconstruction term and a task-specific term, with an optional physical constraint penalty (e.g., smoothness of the filter in frequency space) to prevent overfitting.

**Hypotheses cles :**

- Groundwater level signals are locally stationary within the STFT window (365 days), allowing for meaningful frequency decomposition.
- The spectral characteristics of pumping and climatic signals are sufficiently distinct in at least one frequency band to allow for separation, even if they overlap in others.
- The relationship between the input signal's spectral envelope and the optimal filter is learnable and generalizes across different wells with similar hydrogeological characteristics.
- Groundwater level data is available at a daily temporal resolution for a minimum period of 5 years to capture inter-annual climate variability.

**Inconnues identifiees :**

- We do not know the minimum spectral separation required for the AFFL to outperform fixed filters.
- We do not know whether the learned filters will be physically interpretable to hydrogeologists (e.g., whether they correspond to known aquifer response functions).
- We do not know the extent to which the AFFL's performance degrades with highly non-stationary signals caused by extreme events (e.g., droughts, floods).

### 1.4 Conditions aux limites

- **The temporal resolution of groundwater level data must be at least daily (Δt ≤ 1 day).** — Lower temporal resolution (e.g., weekly) would alias high-frequency pumping signals and prevent the AFFL from learning meaningful frequency filters.
- **The total observation period must be at least 5 years.** — This ensures that at least 5 full seasonal cycles are captured, allowing the STFT to resolve inter-annual climate variability and long-term pumping trends.
- **The spectral overlap ratio (SOR) between pumping and climatic signals must be less than 0.95.** — At SOR > 0.95, the signals are nearly identical in frequency content, making separation theoretically impossible regardless of the method used.
- **The aquifer system must not be subject to abrupt, non-stationary changes (e.g., well drilling, aquifer recharge events) that alter the signal's frequency content over timescales shorter than the STFT window.** — Such abrupt changes would violate the local stationarity assumption and degrade the performance of the frequency-domain filtering.

### 1.5 Cadre theorique

Signal Processing Theory (Fourier Analysis, Adaptive Filtering) integrated with Deep Learning (End-to-End Learning, Representation Learning)

## 2. Etat de l'art et positionnement

### 2.1 Travaux les plus proches

- **[2023] Adaptive Frequency Filters As Efficient Global Token Mixers** — [10.1109/ICCV51070.2023.00556](https://doi.org/10.1109/ICCV51070.2023.00556)
  - Similarite: related
  - Difference cle: This paper applies adaptive frequency filters as global token mixers in vision transformers, but does not extend the concept to 1D hydrological time series or groundwater signal separation.
- **[2024] FilterNet: Harnessing Frequency Filters for Time Series Forecasting** — [10.48550/arXiv.2411.01623](https://doi.org/10.48550/arXiv.2411.01623)
  - Similarite: related
  - Difference cle: FilterNet uses learnable frequency filters for time series forecasting, but focuses on general forecasting accuracy rather than the specific task of separating anthropogenic and climatic signals in groundwater data.
- **[2025] Affirm: Interactive Mamba with Adaptive Fourier Filters for Long-term Time Series Forecasting** — [10.1609/aaai.v39i20.35463](https://doi.org/10.1609/aaai.v39i20.35463)
  - Similarite: related
  - Difference cle: Affirm uses adaptive Fourier filters for time series forecasting, but does not address the hydrological application of separating overlapping signal components for tasks like trend detection or recharge estimation.

### 2.2 Base de preuves

- **[2023] Adaptive Frequency Filters As Efficient Global Token Mixers** — [10.1109/ICCV51070.2023.00556](https://doi.org/10.1109/ICCV51070.2023.00556)
  - Type: analogous | Citations: 117
  - Demonstrates that adaptive frequency filters can be learned in the Fourier domain for vision tasks, providing the core methodological principle that could be transferred to 1D hydrological signals.
- **[2024] FilterNet: Harnessing Frequency Filters for Time Series Forecasting** — [10.48550/arXiv.2411.01623](https://doi.org/10.48550/arXiv.2411.01623)
  - Type: indirect | Citations: 123
  - Shows that learnable frequency filters can be applied to time series data, supporting the feasibility of the proposed transfer to 1D signals.
- **[2025] Affirm: Interactive Mamba with Adaptive Fourier Filters for Long-term Time Series Forecasting** — [10.1609/aaai.v39i20.35463](https://doi.org/10.1609/aaai.v39i20.35463)
  - Type: indirect | Citations: 29
  - Demonstrates the effectiveness of adaptive Fourier filters in time series forecasting, supporting the idea that such filters can handle complex temporal patterns.
- **[2024] Frequency Adaptive Normalization For Non-stationary Time Series Forecasting** — [10.48550/arXiv.2409.20371](https://doi.org/10.48550/arXiv.2409.20371)
  - Type: indirect | Citations: 77
  - Shows that frequency-adaptive methods can handle non-stationarity in time series, which is relevant to groundwater signals that exhibit trends and seasonal patterns.
- **[2024] Predicting groundwater level using traditional and deep machine learning algorithms** — [10.3389/fenvs.2024.1291327](https://doi.org/10.3389/fenvs.2024.1291327)
  - Type: indirect | Citations: 83
  - Confirms that deep learning methods, particularly CNNs, are effective for groundwater level prediction, supporting the feasibility of deep learning approaches in this domain.
- **[2024] Interpretable baseflow segmentation and prediction based on numerical experiments and deep learning.** — [10.1016/j.jenvman.2024.121089](https://doi.org/10.1016/j.jenvman.2024.121089)
  - Type: indirect | Citations: 21
  - Addresses the challenge of separating hydrological signal components (baseflow) and uses deep learning, which is directly relevant to the proposed separation of anthropogenic and climatic signals.

### 2.3 Contre-preuves et limitations connues

- **[minor] Delving Deeper into Anti-Aliasing in ConvNets**
  - This paper focuses on adaptive filtering for anti-aliasing in images, which is a different objective than signal separation. It does not contradict the hypothesis but highlights that adaptive filtering in vision is often for feature preservation, not component separation.

### 2.4 Evaluation de nouveaute

- **Score**: 0.7/1.0
- **Verdict**: novel

## 3. Predictions falsifiables

| # | Prediction | Borne quantitative | Methode | H0 | Test statistique |
|---|-----------|-------------------|---------|-----|-----------------|
| 1 | The AFFL will achieve a lower RMSE in separating pumping signals from climatic signals compared to a standard LSTM baseline and a wavelet-based fixed filter. | AFFL will achieve a mean RMSE reduction of 15-25% (e.g., from 0.40 m to 0.30-0.34 m) on synthetic test data with a spectral overlap ratio of 0.5 and SNR of 20 dB. | Synthetic time series generation with known pumping and climatic components, followed by signal reconstruction and RMSE calculation. | H0: The mean RMSE of the AFFL is not significantly different from the mean RMSE of the LSTM baseline or the wavelet filter. | Paired t-test (two-tailed) with alpha=0.05 and power=0.80, requiring n=25 independent synthetic time series per condition. |
| 2 | The AFFL will maintain a trend estimation error (TEE) of less than 0.05 m/year even when the pumping signal spectrally overlaps with the seasonal climatic signal (SOR=0.8). | TEE < 0.05 m/year for SOR up to 0.8, compared to TEE > 0.15 m/year for the wavelet filter under the same conditions. | Estimate linear trend from the reconstructed pumping signal and compare to the known true trend in synthetic data. | H0: The TEE of the AFFL is not significantly lower than the TEE of the wavelet filter at SOR=0.8. | Mann-Whitney U test (one-tailed) with alpha=0.05, power=0.80, requiring n=30 independent simulations per condition. |
| 3 | The AFFL's learned filters will show consistent spectral patterns across wells with similar hydrogeological characteristics (e.g., same aquifer type). | The mean pairwise cosine similarity of learned filter weights will be > 0.7 for wells within the same aquifer type, and < 0.4 for wells from different aquifer types. | Train AFFL on data from multiple wells, extract the learned filter weights, and compute pairwise cosine similarity matrices. | H0: The mean pairwise cosine similarity of filter weights is not significantly different between same-aquifer and different-aquifer pairs. | Permutation test (10,000 permutations) with alpha=0.05, requiring a minimum of 10 wells per aquifer type. |

## 4. Protocole experimental

**Timeline globale**: 10-16 months
**Budget global**: €25k-120k

### 4.1 Phase 1 — In Silico Validation

**Objectif**: To rapidly falsify the core hypothesis using synthetic data, specifically testing whether the AFFL mechanism can separate spectrally overlapping signals under controlled conditions, and to identify the minimum SOR threshold where the approach fails.

**Methodologie**: 1. **Synthetic Data Generation**: Build a Python-based generator (using NumPy/SciPy) that creates synthetic groundwater level time series with known pumping (e.g., periodic with trend) and climatic (e.g., seasonal + noise) components. Vary the SOR (0.0-0.95) and SNR (5-30 dB) systematically. Generate 100 independent time series per condition (length: 5 years, daily resolution).
2. **Baseline Implementation**: Implement three baselines in PyTorch: (a) a fixed wavelet filter (using PyWavelets, e.g., 'db4' with soft thresholding), (b) a standard LSTM autoencoder for signal separation, and (c) a simple linear regression model on raw time series.
3. **AFFL Implementation**: Implement the AFFL as described: STFT (window=365, hop=30) -> learnable 1D convolutional filter on the frequency axis -> inverse STFT -> regression head. Train end-to-end with a combined loss (reconstruction + task-specific + smoothness penalty).
4. **Evaluation**: For each condition, compute the RMSE between the reconstructed pumping signal and the true pumping signal (SSA). Compute TEE and REE from the reconstructed signals. Run paired t-tests (AFFL vs. each baseline) for each condition.
5. **Ablation Study**: Test the AFFL without the physical constraint penalty and with different filter sizes to understand sensitivity.

- Cout: €500-2000
- Duree: 4-6 weeks
- Equipement: Standard workstation (CPU/GPU)
- Logiciels: Python 3.10+, PyTorch 2.0+, NumPy, SciPy, PyWavelets, Matplotlib, Jupyter Notebook

**Criteres de succes :**

- SSA improvement over baselines: AFFL achieves a mean RMSE reduction of ≥15% compared to the best baseline (LSTM or wavelet) at SOR=0.5 and SNR=20 dB, with p<0.05 in a paired t-test.
- TEE at high SOR: AFFL achieves TEE < 0.05 m/year at SOR=0.8, while the wavelet filter has TEE > 0.15 m/year under the same conditions.
- Minimum viable SOR: Identify the SOR value (SOR_min) below which AFFL's performance degrades to within 5% of the LSTM baseline. This SOR_min must be > 0.7 for the hypothesis to hold.

- **GO**: All three success criteria are met (SSA improvement, TEE at high SOR, SOR_min > 0.7).
- **NO-GO**: AFFL fails to beat the LSTM baseline at SOR=0.5 (criterion 1 not met) OR SOR_min < 0.5, indicating the method is too fragile.
- **PIVOT**: If AFFL works but only with the physical constraint penalty, pivot to a 'physics-informed' AFFL variant. If AFFL fails on synthetic data, pivot to a different deep learning architecture (e.g., a transformer-based filter).

### 4.2 Phase 2 — Minimal Experimental Validation

**Objectif**: To validate the AFFL on a small set of real-world groundwater data, confirming that the learned filters are physically interpretable and that the performance gains observed in silico translate to real data with known or partially known signal components.

**Methodologie**: 1. **Data Acquisition**: Obtain daily groundwater level data and corresponding pumping records for 3-5 wells from a single, well-characterized aquifer (e.g., a USGS site or a European groundwater observatory). Data must span at least 5 years.
2. **Signal Separation**: Apply the trained AFFL model (from Phase 1) to the real data. Also apply the wavelet filter and LSTM baseline.
3. **Validation of Separated Signals**: Compare the AFFL's separated pumping signal to the actual pumping records (if available). If pumping records are not available, compare the 'climatic' component to a known recharge model (e.g., a water balance model) or to precipitation/evapotranspiration data.
4. **Interpretability Analysis**: Extract the learned filter weights from the AFFL for each well. Compute the pairwise cosine similarity of filters from wells in the same aquifer. Visualize the filters to see if they resemble known aquifer response functions (e.g., exponential decay or sinusoidal patterns).
5. **Trend Comparison**: Estimate the long-term trend from the AFFL's separated pumping signal and compare it to the trend from the raw data and from the wavelet filter.

- Cout: €2k-10k
- Duree: 1-2 months
- Equipement: Standard workstation (GPU recommended)
- Logiciels: Python, PyTorch, Pandas, GeoPandas (for spatial data), Matplotlib

**Criteres de succes :**

- Filter interpretability: The mean pairwise cosine similarity of learned filter weights is > 0.7 for wells within the same aquifer type, and < 0.4 for wells from different aquifer types (if multiple types are available).
- Signal separation quality on real data: The AFFL's separated pumping signal shows a correlation of > 0.7 with the observed pumping records (if available) OR the AFFL's separated climatic component shows a correlation of > 0.7 with a water balance model's recharge estimate.
- Trend estimation improvement: The AFFL's trend estimate is closer to the known/estimated true trend than the wavelet filter's estimate, with an improvement of > 20% in absolute error.

- **GO**: All three success criteria are met (filter interpretability, signal separation quality, trend improvement).
- **NO-GO**: The filters are not interpretable (cosine similarity < 0.5 within same aquifer) OR the AFFL fails to separate signals on real data (correlation < 0.5 with reference).
- **PIVOT**: If the filters are interpretable but the separation quality is poor, pivot to a fine-tuning approach (retrain the AFFL on a small portion of the real data). If the filters are not interpretable, pivot to a 'physics-informed' loss that explicitly penalizes non-physical filter shapes.

### 4.3 Phase 3 — Full Experimental Protocol

**Objectif**: To rigorously validate the AFFL against all baselines on a large, diverse dataset of real-world groundwater wells, producing publication-ready results that test all three falsifiable predictions with statistical power.

**Methodologie**: 1. **Large-Scale Data Collection**: Assemble a dataset of 100+ wells from diverse hydrogeological settings (e.g., alluvial, karst, confined/unconfined aquifers) with daily groundwater level data for at least 5 years. Include metadata on aquifer type, pumping rates (if available), and climate data (precipitation, ET).
2. **Preprocessing Pipeline**: Develop a robust preprocessing pipeline to handle missing data (interpolation), outliers, and non-stationary events (e.g., detect and flag abrupt changes).
3. **Model Training and Tuning**: Train the AFFL and all baselines (wavelet, EMD, LSTM, linear regression) on a training set (70% of wells), validate on a validation set (15%), and test on a held-out test set (15%). Use a systematic hyperparameter search (e.g., Optuna) for each model.
4. **Statistical Testing**: For Prediction 1 (SSA improvement), use a paired t-test on the test set wells, with a pre-specified effect size (15% RMSE reduction) and power (0.80). For Prediction 2 (TEE at SOR=0.8), use a Mann-Whitney U test on a subset of wells with high SOR. For Prediction 3 (filter consistency), use a permutation test on the cosine similarity matrix.
5. **Robustness Analysis**: Test the AFFL's performance under different boundary conditions (e.g., weekly data, 3-year records) to assess its limitations. Perform an ablation study to isolate the contribution of each component (STFT, learnable filter, task head).
6. **Publication Output**: Prepare a manuscript with full methodological details, results, and code/data availability.

- Cout: €15k-100k+
- Duree: 6-12 months
- Equipement: GPU cluster or high-performance workstation (e.g., 4x NVIDIA A100 or equivalent)
- Logiciels: Python, PyTorch, Optuna (hyperparameter tuning), DVC (data version control), Weights & Biases (experiment tracking), LaTeX (for manuscript)

**Criteres de succes :**

- Prediction 1: SSA improvement: AFFL achieves a mean RMSE reduction of 15-25% compared to the best baseline (LSTM or wavelet) on the held-out test set, with p<0.05 in a paired t-test.
- Prediction 2: TEE at high SOR: AFFL maintains TEE < 0.05 m/year for wells with SOR up to 0.8, while the wavelet filter has TEE > 0.15 m/year under the same conditions, with p<0.05 in a Mann-Whitney U test.
- Prediction 3: Filter consistency: Mean pairwise cosine similarity of learned filters > 0.7 for wells within the same aquifer type, and < 0.4 for wells from different aquifer types, with p<0.05 in a permutation test.

- **GO**: All three predictions are confirmed with statistical significance. The AFFL demonstrates robust performance across diverse aquifer types and boundary conditions.
- **NO-GO**: Any of the three predictions fails (e.g., AFFL does not beat baselines, or filters are not consistent). This would falsify the core hypothesis.
- **PIVOT**: If the AFFL works but the filters are not physically interpretable, pivot to a 'physics-informed' variant and test if it retains performance. If the AFFL fails on non-stationary data, pivot to a hybrid approach (e.g., AFFL + change-point detection).

### 4.4 Quick start : comment demarrer aujourd'hui

- **Peut demarrer maintenant**: Oui
- **Premiere action**: Create a new Python virtual environment and install PyTorch, NumPy, SciPy, PyWavelets, and Matplotlib. Then, write a script to generate 100 synthetic time series with known pumping and climatic components, with SOR=0.5 and SNR=20 dB, and save them as NumPy arrays.
- **Outils**: Python 3.10+, PyTorch, NumPy, SciPy, PyWavelets
- **Donnees ouvertes**: Synthetic data (self-generated), USGS NWIS (for later phases), European Groundwater Data Portal (for later phases)

## 5. Analyse d'impact

### 5.1 Impact scientifique

Novelty score: 0.7/1.0 (novel)

### 5.2 Applications industrielles et marche

**Score industriel**: 6.5/10

**Forces :**
- Marche de niche mais reel : le marche mondial de la gestion des ressources en eau (smart water management) est estime a ~$20B d'ici 2027 (croissance CAGR ~8%). Les agences gouvernementales (USGS, BRGM, BGS), les bureaux d'etudes (Arcadis, WSP, Ramboll) et les utilities payent deja pour des modeles de prevision et de separation de signaux hydrogeologiques. Un outil qui ameliore la separation pompage/climat repond a un besoin reglementaire (directive-cadre europeenne sur l'eau) et de gestion de la secheresse.
- Avantage competitif distinctif : les filtres adaptatifs appris (AFFL) offrent une robustesse au chevauchement spectral (SOR > 0.7) que les methodes classiques (wavelets, filtres de Kalman) ne fournissent pas. Cela permet de quantifier plus precisement l'impact anthropique sur les aquiferes, un argument de vente direct pour les etudes d'impact environnemental et les litiges sur les droits d'eau.
- Le protocole en 3 phases est un excellent outil de de-risking : le GO/NO-GO a chaque phase permet de limiter l'investissement initial (€500-2000 en Phase 1) avant de commiter des ressources plus importantes. Cela correspond a une approche de developpement agile et frugale, compatible avec un budget R&D de PME.

**Faiblesses :**
- Marche adressable trop etroit pour un ROI massif : la cible principale (hydrogeologues) est un marche de niche. Le volume de licences logicielles vendables est faible compare a des domaines comme la sante ou la finance. Le chiffre d'affaires potentiel (probablement < €5M/an) ne justifie pas un investissement R&D lourd (>€500k) sans une strategie de diversification (ex: appliquer AFFL a d'autres signaux geophysiques ou a la sismique).
- Barriere a l'adoption : les hydrogeologues utilisent des outils matures (MODFLOW, FEFLOW) et sont conservateurs. Convaincre les regulateurs et les bureaux d'etudes de faire confiance a un 'deep learning' non interpretable est un obstacle majeur. La Phase 2 exige une interpretabilite (cosine similarity > 0.5), mais cela ne garantit pas que les clients accepteront une 'boite noire' dans leurs rapports reglementaires.
- Le protocole ne mentionne pas la gestion des donnees manquantes ou irregulieres, frequentes dans les series temporelles de puits. Si l'AFFL echoue sur des donnees reelles incompletes (trous, erreurs de mesure), la commercialisation sera bloque. Le budget Phase 3 (€100k) est sous-estime pour couvrir le nettoyage et la normalisation de donnees multi-sources.

**Recommandation**: Ne pas lancer un produit autonome. Au lieu de cela, developper l'AFFL comme un module 'plug-in' pour les suites logicielles existantes (ex: plug-in pour QGIS ou pour les outils de traitement de signal de MathWorks). Cibler les grands groupes d'ingenierie (Arcadis, AECOM) comme partenaires de co-developpement, en leur offrant une exclusivite temporaire sur un bassin versant pilote. L'objectif est de vendre une licence de module (€20k-50k par an) plutot qu'un logiciel complet. Le ROI est modeste mais le risque est faible. Si la Phase 1 et 2 reussissent, lever des fonds pour une application verticale en gestion de crise secheresse (marche public).

### 5.3 Opportunites de financement

| Programme | Agence | Fit | Budget type | Taux succes |
|-----------|--------|-----|-------------|-------------|
| ANR Jeunes Chercheuses et Jeunes Chercheurs (JCJC) 2026 | Agence Nationale de la Recherche (France) | 0.85 | 150 000 € - 300 000 € (pour 36-48 mois) | ~12-15% (taux de succes historiquement bas, mais competitif pour des projets methodologiques forts) |
| ERC Starting Grant 2026 | European Research Council (Commission Europeenne) | 0.7 | 1 500 000 € - 2 000 000 € (pour 60 mois) | ~14-15% (taux de succes extremement competitif, base sur l'excellence scientifique uniquement) |
| Horizon Europe - Cluster 6 (Food, Bioeconomy, Natural Resources, Agriculture and Environment) - Call: 'Digital solutions for water management' (HORIZON-CL6-2025-WATER-01) | Commission Europeenne (DG RTD / CINEA) | 0.6 | 4 000 000 € - 6 000 000 € (pour 36-48 mois, projets collaboratifs multi-acteurs) | ~20-25% (taux de succes plus eleve que l'ERC mais projets en consortium) |

- **ANR Jeunes Chercheuses et Jeunes Chercheurs (JCJC) 2026** (Agence Nationale de la Recherche (France)): Ce projet est un excellent fit pour le programme JCJC car il est porte par un chercheur junior, il est fondamentalement methodologique (Deep Learning applique a l'hydrogeologie) et il propose un protocole experimental rigoureux. Le budget demande (25-120k) est inferieur au plafond, ce qui est un avantage, mais il faudra justifier la duree (10-16 mois) par rapport a la duree standard de 36-48 mois. Le caractere novateur de l'AFFL repond a l'objectif du programme de soutenir des recherches originales et a risque.
- **ERC Starting Grant 2026** (European Research Council (Commission Europeenne)): L'originalite de l'hypothese et la rigueur du protocole de falsification sont des arguments forts pour l'ERC. Cependant, le projet actuel est trop restreint en termes d'ambition et de duree pour un ERC Starting Grant. Pour etre fundable, il faudrait elargir la portee : par exemple, generaliser la methode AFFL a d'autres signaux geophysiques (sismiques, deformation de surface) ou developper une theorie plus large sur l'apprentissage de filtres adaptatifs pour les series temporelles environnementales. Le budget de 120k est un obstacle, car l'ERC attend un projet a plus grande echelle.
- **Horizon Europe - Cluster 6 (Food, Bioeconomy, Natural Resources, Agriculture and Environment) - Call: 'Digital solutions for water management' (HORIZON-CL6-2025-WATER-01)** (Commission Europeenne (DG RTD / CINEA)): Ce programme correspond si l'objectif est de deployer l'outil a grande echelle pour la gestion operationnelle des ressources en eau. L'hypothese AFFL-GW serait un work package methodologique au sein d'un projet plus large. Le budget actuel (120k) est insuffisant pour ce type de projet, et le consortium devrait inclure des PMEs de developpement logiciel, des agences de l'eau (ex: BRGM, agences de bassin) et des universites. Le narratif devrait etre reoriente de 'preuve de concept' vers 'solution operationnelle pour la resilience a la secheresse'.

**Recommandation financement**: Positionner ce projet comme une preuve de concept methodologique ambitieuse. Pour maximiser les chances, cibler en priorite un financement individuel (ANR JCJC ou ERC Starting Grant) en mettant en avant l'originalite fondamentale et la rigueur du protocole. Si l'objectif est un financement collaboratif, il est imperatif de desormais identifier un partenaire applicatif (agence de l'eau, BRGM) et de reorienter le narratif vers la gestion de la ressource en contexte de stress hydrique, ce qui correspond aux priorites de l'Europe. Le budget de 120k est trop faible pour un projet collaboratif europeen, mais ideal pour un projet individuel.

## 6. Panel Review Summary

| Reviewer | Score | Verdict | Point cle |
|----------|-------|---------|-----------|
| methodologist | 6.5/10 | weak_accept | The three-phase design with explicit GO/NO-GO/PIVOT criteria at each phase is exemplary for staged falsification, preventing premature investment in a flawed approach while allowing for adaptive refinement. |
| domain_expert | 6.8/10 | weak_accept | The hypothesis correctly identifies a critical limitation in standard signal processing for hydrogeology: the non-stationary nature of groundwater signals and the spectral overlap between anthropogenic and climatic forcings. Framing the problem as a learnable, task-conditional filtering operation in the STFT domain is a theoretically sound and modern approach, directly extending the proven concept of adaptive spectral filtering from computer vision (e.g., Adaptive Frequency Filters) to a novel 1D geophysical domain. |
| contrarian | 4.0/10 | weak_reject | The explicit acknowledgment of spectral overlap as a core challenge is a step up from naive frequency-domain filtering approaches, which typically assume disjoint spectral supports. |
| industrialist | 6.5/10 | weak_accept | Marche de niche mais reel : le marche mondial de la gestion des ressources en eau (smart water management) est estime a ~$20B d'ici 2027 (croissance CAGR ~8%). Les agences gouvernementales (USGS, BRGM, BGS), les bureaux d'etudes (Arcadis, WSP, Ramboll) et les utilities payent deja pour des modeles de prevision et de separation de signaux hydrogeologiques. Un outil qui ameliore la separation pompage/climat repond a un besoin reglementaire (directive-cadre europeenne sur l'eau) et de gestion de la secheresse. |
| funding_strategist | 7.2/10 | accept | Originalite methodologique : l'integration d'une couche de filtrage frequentiel apprenable (AFFL) dans un reseau de neurones pour la separation de signaux hydrogeologiques est une contribution theorique nouvelle, au-dela des approches standard de traitement du signal (filtres fixes) ou de Deep Learning pur (LSTM). Le concept de 'spectral overlap robustness' cible une limitation reelle et documentee des methodes actuelles. |

**Consensus score**: 6.2/10
**Verdict final**: publish_brief

### 6.1 Consensus

- Tous les reviewers reconnaissent la pertinence du problème (chevauchement spectral pompage/climat) et l'originalité de l'approche AFFL comme une avancée méthodologique crédible par rapport aux filtres fixes.
- Le protocole en 3 phases avec critères GO/NO-GO quantitatifs est unanimement salué comme un exemple de rigueur scientifique et de falsifiabilité, réduisant le risque perçu d'échec.
- Le principal point de convergence critique est le besoin de clarifier la définition opérationnelle du SOR (Spectral Overlap Ratio) pour les données réelles, et de justifier physiquement la fenêtre STFT de 365 jours face à la non-stationnarité multi-échelle des signaux hydrogéologiques.

### 6.2 Points de desaccord

- Le Contrarian (confiance 0.85) estime que la fenêtre STFT de 365 jours est un choix fatal qui rend le problème mal posé pour les signaux inter-annuels (ENSO, PDO), tandis que le Domain Expert et le Methodologist considèrent ce choix comme un hyperparamètre à justifier ou à adapter, pas une fatalité.
- Le Contrarian juge la supériorité de l'AFFL non prouvée face à des baselines physiques simples (filtre passe-bande dérivé d'une fonction de réponse d'aquifère), alors que le Funding Strategist et l'Industrialist voient dans l'AFFL un avantage compétitif distinctif pour les SOR élevés, sans exiger cette comparaison comme prérequis absolu.
- Le Methodologist et le Contrarian s'opposent sur la validité statistique : le premier demande une prise en compte de l'autocorrélation spatiale (risque de pseudo-réplication), le second affirme que les prédictions (15-25% RMSE) sont infalsifiables sans spécification de la géométrie spectrale du chevauchement.

### 6.3 Critical path

Le facteur le plus déterminant est la démonstration que l'AFFL offre un avantage réel et reproductible sur des données réelles, ce qui nécessite de résoudre le problème de la définition opérationnelle du SOR et de la fenêtre temporelle. Si le SOR ne peut pas être calculé de manière robuste sur des chroniques non-stationnaires, ou si la fenêtre de 365 jours ne capture pas les signaux pertinents, l'hypothèse s'effondre en Phase 3. La validation sur données synthétiques (Phase 1) est nécessaire mais insuffisante ; le succès dépend de la capacité à démontrer une séparation physiquement interprétable et statistiquement significative sur des puits réels, en surmontant les défis de l'autocorrélation spatiale et de la qualité des données.

**Recommandation finale**: Le panel recommande la publication du brief, mais avec une mise en garde explicite sur les risques méthodologiques identifiés. L'hypothèse est suffisamment originale et le protocole suffisamment rigoureux pour justifier une investigation approfondie. Cependant, le brief doit être amendé pour inclure une spécification détaillée de la mesure du SOR sur données réelles, une justification physique de la fenêtre STFT (ou une proposition de fenêtre multi-échelle), et un plan d'analyse statistique tenant compte de l'autocorrélation spatiale. La comparaison avec des baselines physiques simples (filtre de Kalman, réponse d'aquifère) est fortement recommandée pour renforcer la crédibilité de la supériorité de l'AFFL. Le potentiel de financement est réel (ANR JCJC en priorité), mais le positionnement applicatif doit être clarifié pour maximiser l'impact.

## 7. Gap Manifest residuel

### 7.1 Data gaps

- No paper in the provided list explicitly applies adaptive frequency filtering to separate anthropogenic and climatic signals in groundwater time series, indicating a clear gap in the literature.
- There is a lack of labeled groundwater datasets with known signal components (e.g., pumping events vs. climate trends) that would be needed to train and evaluate such a model.

### 7.2 Competence gaps

- Phase 1: Deep learning engineering (PyTorch)
- Phase 1: Signal processing (STFT, wavelets)
- Phase 1: Statistical testing (hypothesis testing, power analysis)
- Phase 2: Hydrogeology (data interpretation)
- Phase 2: Deep learning (model application)
- Phase 2: Time series analysis
- Phase 3: Hydrogeology (domain expertise)
- Phase 3: Advanced deep learning (model design, training at scale)
- Phase 3: Statistical analysis (power analysis, mixed-effects models)
- Phase 3: Scientific writing

### 7.3 Epistemic gaps

- We do not know the minimum spectral separation required for the AFFL to outperform fixed filters.
- We do not know whether the learned filters will be physically interpretable to hydrogeologists (e.g., whether they correspond to known aquifer response functions).
- We do not know the extent to which the AFFL's performance degrades with highly non-stationary signals caused by extreme events (e.g., droughts, floods).

## References

[1] Zhipeng Huang, Zhizheng Zhang, et al. (2023). *Adaptive Frequency Filters As Efficient Global Token Mixers*. DOI: [10.1109/ICCV51070.2023.00556](https://doi.org/10.1109/ICCV51070.2023.00556)
[2] Kun Yi, Jingru Fei, et al. (2024). *FilterNet: Harnessing Frequency Filters for Time Series Forecasting*. DOI: [10.48550/arXiv.2411.01623](https://doi.org/10.48550/arXiv.2411.01623)
[3] Yuhan Wu, Xiyu Meng, et al. (2025). *Affirm: Interactive Mamba with Adaptive Fourier Filters for Long-term Time Series Forecasting*. DOI: [10.1609/aaai.v39i20.35463](https://doi.org/10.1609/aaai.v39i20.35463)
[4] Weiwei Ye, Songgaojun Deng, et al. (2024). *Frequency Adaptive Normalization For Non-stationary Time Series Forecasting*. DOI: [10.48550/arXiv.2409.20371](https://doi.org/10.48550/arXiv.2409.20371)
[5] Fan Feng, Hamzeh Ghorbani, et al. (2024). *Predicting groundwater level using traditional and deep machine learning algorithms*. DOI: [10.3389/fenvs.2024.1291327](https://doi.org/10.3389/fenvs.2024.1291327)
[6] Qiying Yu, Chen Shi, et al. (2024). *Interpretable baseflow segmentation and prediction based on numerical experiments and deep learning.*. DOI: [10.1016/j.jenvman.2024.121089](https://doi.org/10.1016/j.jenvman.2024.121089)
[7] Xueyan Zou, Fanyi Xiao, et al. (2020). *Delving Deeper into Anti-Aliasing in ConvNets*. DOI: [10.1007/s11263-022-01672-y](https://doi.org/10.1007/s11263-022-01672-y)

## Annexes

### A. Detailed Reviewer Reports

#### Methodologist

- **Score**: 6.5/10 | **Verdict**: weak_accept | **Confidence**: 0.8
- **Strengths**: The three-phase design with explicit GO/NO-GO/PIVOT criteria at each phase is exemplary for staged falsification, preventing premature investment in a flawed approach while allowing for adaptive refinement.; The use of synthetic data with known ground truth (Phase 1) is a powerful tool for isolating the core mechanism's efficacy under controlled conditions, and the systematic variation of SOR and SNR is methodologically sound for mapping the failure boundary.; The inclusion of an interpretability analysis (cosine similarity of learned filters) in Phase 2 is a strong, non-obvious check that bridges the gap between purely predictive performance and physical plausibility, which is crucial for acceptance in the geosciences field.
- **Weaknesses**: The protocol lacks a pre-registered analysis plan for the 'ablation study' and 'sensitivity analysis' mentioned in Phase 1. Without pre-specifying which perturbations are considered 'acceptable' vs. 'fatal', these explorations risk becoming post-hoc rationalizations of the results, undermining the confirmatory value of the study.; The definition and estimation of the 'Spectral Overlap Ratio (SOR)' from real-world data (Phase 3) is a critical, unresolved methodological gap. The protocol does not specify how SOR will be calculated for field data, which introduces a high risk of measurement bias and makes the selection of a 'high SOR' subset for Prediction 2 arbitrary and non-reproducible.; The statistical power analysis is superficial. The protocol mentions 'power (0.80)' for Prediction 1 but does not specify the expected variance in RMSE reduction across the 100+ wells. Given the known heterogeneity in hydrogeological systems, the assumed effect size (15-25%) may be overly optimistic, and the test may be severely underpowered without a more detailed variance model.
- **Questions**: For Phase 3, how will the Spectral Overlap Ratio (SOR) be operationally defined and calculated from real, non-stationary groundwater level time series? What specific algorithm and parameters will be used to ensure this metric is consistent and reproducible across different wells and aquifer types?; The protocol proposes a paired t-test for Prediction 1. Given that the wells in the test set are likely to be spatially autocorrelated (e.g., wells in the same aquifer are not independent), how will you account for this spatial dependency to avoid pseudo-replication and inflated Type I error rates?; In Phase 2, the success criterion for 'Signal separation quality' relies on a correlation > 0.7 with observed pumping records. What is the statistical power of this test given the expected noise in real-world pumping data, and how will you handle wells where pumping records are incomplete or unreliable?
- **Recommendation**: This is a well-structured protocol with a strong emphasis on staged falsification and physical interpretability. However, before proceeding to Phase 3, significant work is needed to operationalize the SOR metric for real-world data and to develop a more rigorous statistical model that accounts for spatial autocorrelation and realistic data variance. I recommend focusing the next revision on these critical methodological details, particularly the pre-registration of the analysis pipeline for the large-scale validation.

#### Domain Expert

- **Score**: 6.8/10 | **Verdict**: weak_accept | **Confidence**: 0.8
- **Strengths**: The hypothesis correctly identifies a critical limitation in standard signal processing for hydrogeology: the non-stationary nature of groundwater signals and the spectral overlap between anthropogenic and climatic forcings. Framing the problem as a learnable, task-conditional filtering operation in the STFT domain is a theoretically sound and modern approach, directly extending the proven concept of adaptive spectral filtering from computer vision (e.g., Adaptive Frequency Filters) to a novel 1D geophysical domain.; The proposed causal chain is methodologically coherent and end-to-end trainable. The inclusion of a task-specific loss term alongside a reconstruction term, with an optional physical smoothness constraint on the filter, is a well-designed regularization strategy that balances data-driven optimization with physical plausibility. This mitigates the risk of learning a filter that is mathematically optimal but physically meaningless.
- **Weaknesses**: The core assumption of local stationarity within a 365-day window is physically fragile. Groundwater systems are often characterized by multi-year memory, delayed responses to recharge, and abrupt changes due to pumping events or well maintenance. A fixed 365-day window may not capture the relevant low-frequency climatic (e.g., ENSO, PDO) or long-term anthropogenic trends, potentially leading to poor separation of the very signals of interest. The choice of window size is a hyperparameter that requires strong physical justification, not just algorithmic convenience.; The hypothesis lacks a clear definition of 'superior separation' and a robust evaluation protocol. How will the separated pumping signal be validated against independent data (e.g., pumping logs, well discharge records)? Without a ground-truth or a well-defined proxy, the task-specific loss (e.g., trend estimation) could be optimized by learning a filter that simply scales the input, rather than performing meaningful spectral separation. The claim of 'spectral overlap robustness' needs a quantitative definition of the overlap conditions tested.; The positioning overstates the novelty slightly. While the application to groundwater is new, the core mechanism (learnable frequency masking) is a mature concept in deep learning. The key differentiator must be the physical interpretability and the specific constraints of hydrological signals. The hypothesis does not adequately address how the learned filters will be mapped to physically meaningful aquifer response functions (e.g., recession constants, storage coefficients) to provide new hydrogeological insight, which is a key promise of the work.
- **Questions**: Given the multi-year autocorrelation of groundwater head, how does the proposed 365-day STFT window handle signals with significant power at periods longer than the window (e.g., decadal climate oscillations)? Does the model rely on the task-specific head to compensate for the information lost by the windowing, or is there a plan for a multi-scale or dilated frequency decomposition?; What is the specific physical mechanism that the learnable filter is expected to discover? For instance, is the hypothesis that pumping signals have a distinct spectral signature (e.g., high-frequency, non-periodic) compared to climatic signals (e.g., seasonal, low-frequency)? If so, what is the expected shape of the optimal filter, and how will this be validated against known aquifer response functions from pumping tests?; How will the model's performance be evaluated in a way that proves 'spectral overlap robustness'? The evaluation must include synthetic datasets with controlled and known spectral overlap between the two signals, where the ground-truth separation is known, to demonstrate that the AFFL genuinely outperforms fixed filters (e.g., high-pass, low-pass, or band-stop) and standard deep learning baselines (e.g., autoencoders) under varying Signal-to-Noise Ratios (SNR) and overlap conditions.
- **Recommendation**: The hypothesis is promising and well-positioned at the intersection of signal processing, deep learning, and hydrology. However, the current formulation requires a more rigorous treatment of the physical timescales involved and a clearer definition of success. I recommend a revision that explicitly addresses the issue of multi-scale non-stationarity, proposes a validation strategy using synthetic data with known spectral overlap, and outlines a pathway for physical interpretation of the learned filters. With these clarifications, the work could make a significant contribution.

#### Contrarian

- **Score**: 4.0/10 | **Verdict**: weak_reject | **Confidence**: 0.85
- **Strengths**: The explicit acknowledgment of spectral overlap as a core challenge is a step up from naive frequency-domain filtering approaches, which typically assume disjoint spectral supports.; The proposed end-to-end training with a physical smoothness constraint on the filter shows awareness of the overfitting risk inherent in learnable spectral masks.
- **Weaknesses**: FAIL REASON #1: The fundamental assumption of local stationarity within a 365-day window is almost certainly violated for the exact signals the method targets. Anthropogenic pumping is often seasonal (agricultural) but with inter-annual variability tied to drought, and climatic signals (ENSO, PDO) have dominant periods at 2-7 years. A 365-day window cannot resolve these low-frequency, non-stationary components; it will smear them into the DC and low-frequency bins, making the 'spectral separation' task ill-posed from the start. The STFT window size is a fatal hyperparameter choice that contradicts the stated goal of separating inter-annual climatic signals.; FAIL REASON #2: The prediction bounds are internally inconsistent and physically implausible. Claiming a 15-25% RMSE reduction at SOR=0.5 is meaningless without specifying the spectral shape of the overlap. If the overlap is in the seasonal band (annual cycle), then a simple annual harmonic removal (a fixed filter) would achieve near-perfect separation, and the AFFL's advantage vanishes. If the overlap is in the broad low-frequency band, then the 365-day STFT window cannot capture it. The 'spectral overlap ratio' is an ill-defined metric that obscures the actual frequency geometry, making the prediction unfalsifiable.; FAIL REASON #3: The method conflates correlation with causation in the learned filters. The claim that filters will be 'physically interpretable' (cosine similarity > 0.7 within aquifer type) is a post-hoc rationalization. The AFFL is trained end-to-end for a task (e.g., trend estimation). It will find any spectral manipulation that reduces loss, including exploiting artifacts of the STFT (e.g., leakage, edge effects) or the reconstruction loss. There is no mechanism in the architecture that forces the filter to isolate the 'pumping' signal specifically; it only minimizes a composite loss. The learned filters are likely to be non-unique, overfit to the synthetic data generation process, and will not transfer to real wells where the pumping signal is not a simple additive component but interacts non-linearly with aquifer storage and recharge.
- **Questions**: Can you demonstrate that the AFFL's advantage persists when the STFT window is increased to 1825 days (5 years) to capture inter-annual variability, and if so, how does the 'per-sample' weighting remain computationally tractable and statistically robust with only a handful of effective time-frequency bins?; What is the specific frequency band (in cycles per year) where the pumping signal is expected to dominate over climate, and why is a learnable filter superior to a fixed, physically-derived band-pass filter (e.g., an aquifer response function) in that band? If the overlap is in the seasonal band, does your method outperform a simple 'remove the annual cycle' baseline?
- **Recommendation**: The authors must first define the spectral geometry of the problem. They need to provide a spectral analysis of real groundwater data showing that pumping and climate signals have distinct, learnable frequency signatures that are not captured by fixed filters. Then, they must reformulate the STFT window to match the temporal scales of the target signals (likely multi-year). Finally, they should compare against a physically-based baseline (e.g., a linear reservoir model with known frequency response) to prove that the 'learned' filter is not just rediscovering a known transfer function. Without these steps, the claimed 'superiority' is an artifact of the synthetic data setup.

#### Industrialist

- **Score**: 6.5/10 | **Verdict**: weak_accept | **Confidence**: 0.65
- **Strengths**: Marche de niche mais reel : le marche mondial de la gestion des ressources en eau (smart water management) est estime a ~$20B d'ici 2027 (croissance CAGR ~8%). Les agences gouvernementales (USGS, BRGM, BGS), les bureaux d'etudes (Arcadis, WSP, Ramboll) et les utilities payent deja pour des modeles de prevision et de separation de signaux hydrogeologiques. Un outil qui ameliore la separation pompage/climat repond a un besoin reglementaire (directive-cadre europeenne sur l'eau) et de gestion de la secheresse.; Avantage competitif distinctif : les filtres adaptatifs appris (AFFL) offrent une robustesse au chevauchement spectral (SOR > 0.7) que les methodes classiques (wavelets, filtres de Kalman) ne fournissent pas. Cela permet de quantifier plus precisement l'impact anthropique sur les aquiferes, un argument de vente direct pour les etudes d'impact environnemental et les litiges sur les droits d'eau.; Le protocole en 3 phases est un excellent outil de de-risking : le GO/NO-GO a chaque phase permet de limiter l'investissement initial (€500-2000 en Phase 1) avant de commiter des ressources plus importantes. Cela correspond a une approche de developpement agile et frugale, compatible avec un budget R&D de PME.
- **Weaknesses**: Marche adressable trop etroit pour un ROI massif : la cible principale (hydrogeologues) est un marche de niche. Le volume de licences logicielles vendables est faible compare a des domaines comme la sante ou la finance. Le chiffre d'affaires potentiel (probablement < €5M/an) ne justifie pas un investissement R&D lourd (>€500k) sans une strategie de diversification (ex: appliquer AFFL a d'autres signaux geophysiques ou a la sismique).; Barriere a l'adoption : les hydrogeologues utilisent des outils matures (MODFLOW, FEFLOW) et sont conservateurs. Convaincre les regulateurs et les bureaux d'etudes de faire confiance a un 'deep learning' non interpretable est un obstacle majeur. La Phase 2 exige une interpretabilite (cosine similarity > 0.5), mais cela ne garantit pas que les clients accepteront une 'boite noire' dans leurs rapports reglementaires.; Le protocole ne mentionne pas la gestion des donnees manquantes ou irregulieres, frequentes dans les series temporelles de puits. Si l'AFFL echoue sur des donnees reelles incompletes (trous, erreurs de mesure), la commercialisation sera bloque. Le budget Phase 3 (€100k) est sous-estime pour couvrir le nettoyage et la normalisation de donnees multi-sources.
- **Questions**: Quel est le cout de revient reel d'une etude pilote pour un client (ex: un syndicat de bassin versant) et quel prix de vente peut-on justifier par rapport a une etude classique avec wavelets ? Un prix > €50k par projet est-il tenable ?; Peut-on breveter l'AFFL ou est-ce que la formulation (couche de filtrage apprise) est deja anticipee par des brevets sur les 'learnable filters' en audio ou en traitement d'image ? L'IP est-elle defensable face a Google DeepMind ou Schlumberger qui ont des departements ML applique aux geosciences ?; Quel est le plan de diversification pour amortir la R&D ? Peut-on appliquer l'AFFL a la separation de signaux en sismique reflechissante, en electromagnetisme ou en surveillance de barrages (auscultation) ? Le marche de la geophysique de subsurface est 10x plus grand que l'hydrogeologie seule.
- **Recommendation**: Ne pas lancer un produit autonome. Au lieu de cela, developper l'AFFL comme un module 'plug-in' pour les suites logicielles existantes (ex: plug-in pour QGIS ou pour les outils de traitement de signal de MathWorks). Cibler les grands groupes d'ingenierie (Arcadis, AECOM) comme partenaires de co-developpement, en leur offrant une exclusivite temporaire sur un bassin versant pilote. L'objectif est de vendre une licence de module (€20k-50k par an) plutot qu'un logiciel complet. Le ROI est modeste mais le risque est faible. Si la Phase 1 et 2 reussissent, lever des fonds pour une application verticale en gestion de crise secheresse (marche public).

#### Funding Strategist

- **Score**: 7.2/10 | **Verdict**: accept | **Confidence**: 0.82
- **Strengths**: Originalite methodologique : l'integration d'une couche de filtrage frequentiel apprenable (AFFL) dans un reseau de neurones pour la separation de signaux hydrogeologiques est une contribution theorique nouvelle, au-dela des approches standard de traitement du signal (filtres fixes) ou de Deep Learning pur (LSTM). Le concept de 'spectral overlap robustness' cible une limitation reelle et documentee des methodes actuelles.; Protocole de validation exemplaire et falsifiable : le design en trois phases (in silico, minimal, full) avec des criteres GO/NO-GO quantitatifs predefinis (SSA, SOR_min, cosine similarity) est exactement le type de rigueur scientifique que les evaluateurs ERC et ANR recherchent. Cela demontre une maturite epistemologique rare pour un projet de cette taille et reduit le risque percu d'echec.; TRL initial faible mais trajectoire claire et rapide : le passage de TRL 2-3 (concept et validation numerique) a TRL 4-5 (validation en environnement reel) en 16 mois est credible et bien planifie, avec un budget modeste permettant une preuve de concept rapide avant un eventuel scale-up.
- **Weaknesses**: Manque de positionnement applicatif concret : l'hypothese reste academique. Pour des programmes comme Horizon Europe (Cluster 6) ou les appels ANR generiques, il manque une articulation explicite avec des cas d'usage operationnels (gestion de la ressource, secheresse, intrusion saline) et des utilisateurs finaux (agences de l'eau, bureaux d'etudes). Le consortium propose est implicite, ce qui est un handicap pour les projets collaboratifs.; Risque de sur-ingenierie et de generalisation : la superiorite de l'AFFL sur des baselines simples (LSTM) est postulee, mais le protocole ne mentionne pas de comparaison avec des methodes de pointe en separation de sources (ex: ICA, NMF, ou des architectures attentionnelles recentes). Un evaluateur pourrait douter que la complexite ajoutee soit justifiee par rapport a des methodes plus simples et eprouvees.; Dependance critique a la qualite des donnees reelles (Phase 2 et 3) : l'interpretabilite des filtres (criteres de cosine similarity) est un objectif ambitieux qui peut etre difficile a atteindre sur des aquiferes complexes. Le plan ne specifie pas clairement la provenance des donnees reelles, leur accessibilite, ou les accords de partage de donnees avec les organismes gestionnaires, ce qui est un risque logistique majeur.
- **Questions**: Quel est le positionnement par rapport aux methodes de separation de sources aveugles (BSS) classiques (ICA, SOBI) qui sont deja utilisees en hydrogeophysique ? Une comparaison avec ces methodes est-elle prevue pour demontrer la superiorite reelle du Deep Learning ?; Quel est l'angle de valorisation ? Le projet vise-t-il un article de recherche fondamentale (fit ERC/ANR JCJC) ou un outil operationnel (fit Horizon Europe/ANR PRCE) ? Cette distinction conditionne le choix du programme et la composition du consortium.; Quelles sont les sources de donnees reelles envisagees pour la Phase 2 et 3 ? Avez-vous deja des accords de principe avec des agences de l'eau ou des services geologiques (ex: BRGM, USGS) pour acceder a des chroniques piezometriques longues et fiables avec des metadonnees sur les pompages ?
- **Recommendation**: Positionner ce projet comme une preuve de concept methodologique ambitieuse. Pour maximiser les chances, cibler en priorite un financement individuel (ANR JCJC ou ERC Starting Grant) en mettant en avant l'originalite fondamentale et la rigueur du protocole. Si l'objectif est un financement collaboratif, il est imperatif de desormais identifier un partenaire applicatif (agence de l'eau, BRGM) et de reorienter le narratif vers la gestion de la ressource en contexte de stress hydrique, ce qui correspond aux priorites de l'Europe. Le budget de 120k est trop faible pour un projet collaboratif europeen, mais ideal pour un projet individuel.

### B. Semantic Scholar Search Queries

- [novelty] `adaptive frequency filtering groundwater time series` — Direct search for the core hypothesis combining adaptive filtering with groundwater signal processing
- [novelty] `learnable filter Fourier domain hydrological signals` — Targets the specific mechanism of learnable frequency-domain filtering applied to hydrology
- [novelty] `deep learning spatiotemporal filter separation pumping climate` — Searches for end-to-end learned filters separating anthropogenic and climatic signals in water level data
- [novelty] `global adaptive filtering layer geoscience` — Checks if the GAFL concept has already been transferred to any geoscience domain
- [evidence] `frequency domain filtering groundwater level decomposition` — Finds existing frequency-domain methods for decomposing groundwater signals into components
- [evidence] `wavelet transform pumping signal separation aquifer` — Identifies current fixed-filter approaches (wavelet) used to separate anthropogenic from natural signals
- [evidence] `Fourier transform hydrological time series trend extraction` — Provides evidence of frequency-domain techniques for trend and seasonal component extraction in hydrology
- [evidence] `convolutional neural network groundwater level prediction` — Shows existing deep learning applications to groundwater data, establishing a baseline for learnable layers
- [cross_domain] `adaptive filtering layer computer vision frequency` — Finds the original GAFL paper and related adaptive filtering work in vision to confirm the source mechanism
- [cross_domain] `learnable spectral filtering signal processing` — Searches for learnable frequency-domain filtering across other signal processing domains beyond vision
- [cross_domain] `domain adaptation frequency filtering time series` — Looks for interdisciplinary precedents of transferring adaptive frequency filtering concepts to time-series data

---

*Generated by SPORE (Systeme de Production d'Opportunites de Recherche par Exploration) on 2026-08-08*