# Knowledge-Augmented Spectral Deconvolution: An Ontology-Based Semantic Normalization Framework for UV-Vis Wastewater Monitoring

## Metadata

- **SPORE ID**: SPR-2026-CD79
- **Domaines**: Biomedical Text Mining and Ontologies x Water Quality Monitoring and Analysis
- **Date de generation**: 2026-08-17
- **Panel consensus score**: 6.5/10
- **Novelty score**: 0.8/1.0
- **Panel verdict**: publish_brief

## Abstract

If a UV-Vis spectral regression model is augmented with a structured chemical ontology encoding interferent co-occurrence and spectral overlap priors (semantic normalization), then its quantitative prediction accuracy for individual analyte concentrations will improve by a statistically significant margin over a strong, regularized baseline, because the ontology constrains the solution space by disambiguating overlapping spectral contributions.

The proposed mechanism involves 4 causal steps: (1) Construction of a quantitative chemical ontology (Phase 0) for a controlled tern -> (2) Encoding of each UV-Vis spectrum into a graph structure where nodes represent de -> (3) A Graph Neural Network (GNN) or a semantic attention layer processes this graph, -> (4) The output layer of the model predicts the concentration of each target analyte,.

Literature grounding on 15 verified references yields a novelty score of 0.8 (novel). A 3-phase experimental protocol (budget: €25k-120k, timeline: 10-18 months) is proposed, starting with in silico validation. A panel of 5 expert reviewers reached a consensus score of 6.5/10.

## 1. Hypothese et mecanisme propose

### 1.1 Formulation formelle

If a UV-Vis spectral regression model is augmented with a structured chemical ontology encoding interferent co-occurrence and spectral overlap priors (semantic normalization), then its quantitative prediction accuracy for individual analyte concentrations will improve by a statistically significant margin over a strong, regularized baseline, because the ontology constrains the solution space by disambiguating overlapping spectral contributions.

### 1.2 Variables

**Variables independantes :**

| Variable | Type | Plage | Unite |
|----------|------|-------|-------|
| Model Augmentation Type | categorical | Baseline (LASSO/Elastic Net) vs. Ontology-Augmented (Graph Neural Network with spectral + ontology embeddings) | N/A |
| Ontology Coverage of Interferents | categorical | Full (all known interferents) vs. Partial (80% of interferents) vs. Novel (interferent excluded from ontology) | N/A |

**Variables dependantes :**

| Variable | Type | Direction attendue | Unite |
|----------|------|-------------------|-------|
| Mean Absolute Percentage Error (MAPE) for individual analyte concentration prediction | continuous | decrease | % |
| Recall of novel interferent detection | continuous | increase | proportion (0-1) |

### 1.3 Chaine causale

1. Step 1: Construction of a quantitative chemical ontology (Phase 0) for a controlled ternary metal mixture (Cu, Zn, Ni) in a buffer, encoding spectral peak positions, molar absorptivities, pH-dependent speciation, and known co-occurrence patterns from industrial discharge data.
1. Step 2: Encoding of each UV-Vis spectrum into a graph structure where nodes represent detected spectral features (peaks/shoulders) and edges represent potential chemical relationships (e.g., known overlap, co-occurrence probability) derived from the ontology.
1. Step 3: A Graph Neural Network (GNN) or a semantic attention layer processes this graph, using the ontology embeddings to weight spectral features. This allows the model to 'disambiguate' a spectral region by attributing its absorbance to the most chemically plausible interferent based on the co-occurrence priors.
1. Step 4: The output layer of the model predicts the concentration of each target analyte, with the ontology-derived weights acting as a regularizer that penalizes chemically implausible concentration combinations (e.g., high Cu and high Ni in a sample from a process known to only use Cu).

**Hypotheses cles :**

- Assumption 1: The spectral signatures of the target analytes and interferents are additive and follow the Beer-Lambert law within the concentration ranges tested (0.1-10 mg/L).
- Assumption 2: The chemical ontology built from external data (e.g., a different industrial site) captures the fundamental chemical properties (e.g., molar absorptivity, pKa) that are transferable across sites, even if the co-occurrence probabilities are site-specific.
- Assumption 3: The baseline model (LASSO/Elastic Net) is optimally tuned, ensuring that any performance gain from the ontology-augmented model is not due to underfitting of the baseline.

**Inconnues identifiees :**

- We do not know if the ontology's co-occurrence priors, built from one industrial site, will be sufficiently generalizable to improve predictions at a different site with different discharge profiles.
- We do not know the minimum size and completeness of the ontology required to achieve a significant improvement over the baseline.
- We do not know if the model can effectively flag a novel interferent (not in the ontology) without a significant drop in accuracy for known analytes.

### 1.4 Conditions aux limites

- **Analyte concentrations must be within the linear range of the Beer-Lambert law (0.1 - 10 mg/L for Cu, Zn, Ni).** — Above 10 mg/L, aggregation or precipitation can occur, violating the additivity assumption of the spectral mixture model. Below 0.1 mg/L, the signal-to-noise ratio of the spectrophotometer is too low for reliable quantification.
- **The pH of the sample matrix must be buffered to within ±0.2 pH units of the calibration pH (e.g., pH 5.0).** — The speciation of metal ions (and thus their UV-Vis spectra) is highly pH-dependent. Uncontrolled pH would introduce a source of variance not captured by the ontology, breaking the causal link between ontology priors and spectral features.
- **The ontology must be constructed from a domain that is chemically analogous to the target domain (e.g., electroplating industry wastewater for a model tested on mining wastewater).** — The core mechanism relies on the transferability of chemical properties (spectra, pKa) and co-occurrence logic. If the ontology is built for a completely unrelated chemical context, the priors will be misleading and could degrade performance.

### 1.5 Cadre theorique

Knowledge-Augmented Machine Learning (KA-ML) and Semantic Sensor Web Theory. The hypothesis posits that integrating structured domain knowledge (ontology) into a learning algorithm acts as a powerful inductive bias, constraining the hypothesis space and improving generalization, particularly in low-data or high-ambiguity regimes.

## 2. Etat de l'art et positionnement

### 2.1 Travaux les plus proches

- **[2019] Simultaneous determination of trace metal ions in industrial wastewater based on UV–vis spectrometry** — [10.1016/J.IJLEO.2018.09.075](https://doi.org/10.1016/J.IJLEO.2018.09.075)
  - Similarite: related
  - Difference cle: This paper addresses spectral overlap in wastewater using signal processing (wavelet transform) but does not use ontology or semantic normalization; the hypothesis proposes a knowledge-augmented approach.
- **[2024] Medical-informed machine learning: integrating prior knowledge into medical decision systems** — [10.1186/s12911-024-02582-4](https://doi.org/10.1186/s12911-024-02582-4)
  - Similarite: analogous
  - Difference cle: This paper integrates prior knowledge into ML for medical decisions, but not for spectral analysis or wastewater monitoring; the hypothesis transfers this concept to a different domain.
- **[2023] Harnessing Prior Knowledge for Explainable Machine Learning: An Overview** — [10.1109/SaTML54575.2023.00038](https://doi.org/10.1109/SaTML54575.2023.00038)
  - Similarite: analogous
  - Difference cle: This paper reviews prior knowledge integration in ML for explainability, but not specifically for spectral deconvolution or using ontologies for chemical interferents.

### 2.2 Base de preuves

- **[2019] Simultaneous determination of trace metal ions in industrial wastewater based on UV–vis spectrometry** — [10.1016/J.IJLEO.2018.09.075](https://doi.org/10.1016/J.IJLEO.2018.09.075)
  - Type: direct | Citations: 18
  - Directly demonstrates the problem of spectral overlap for Cu, Co, and Ni in wastewater and attempts to resolve it with signal processing, which is the core problem the hypothesis aims to solve.
- **[2024] Medical-informed machine learning: integrating prior knowledge into medical decision systems** — [10.1186/s12911-024-02582-4](https://doi.org/10.1186/s12911-024-02582-4)
  - Type: analogous | Citations: 61
  - Supports the core idea of integrating domain knowledge into ML models, which is the paradigm shift proposed in the hypothesis.
- **[2023] Harnessing Prior Knowledge for Explainable Machine Learning: An Overview** — [10.1109/SaTML54575.2023.00038](https://doi.org/10.1109/SaTML54575.2023.00038)
  - Type: analogous | Citations: 23
  - Supports the idea that prior knowledge can improve model performance and explainability, which is a key motivation for the hypothesis.
- **[2011] A Semantic Sensor Web for Environmental Decision Support Applications** — [10.3390/s110908855](https://doi.org/10.3390/s110908855)
  - Type: indirect | Citations: 45
  - Shows the use of semantic web technologies for environmental monitoring, supporting the feasibility of a semantic layer for wastewater data.
- **[2017] Integrating Statistical Machine Learning in a Semantic Sensor Web for Proactive Monitoring and Control** — [10.3390/s17040807](https://doi.org/10.3390/s17040807)
  - Type: indirect | Citations: 25
  - Demonstrates the integration of ML with semantic sensor web for environmental monitoring, a key architectural component of the hypothesis.
- **[2013] Ontology Alignment Architecture for Semantic Sensor Web Integration** — [10.3390/s130912581](https://doi.org/10.3390/s130912581)
  - Type: indirect | Citations: 38
  - Addresses the challenge of integrating heterogeneous ontologies in sensor webs, which is relevant to building a chemical ontology for wastewater.
- **[2024] Identifying compound-protein interactions with knowledge graph embedding of perturbation transcriptomics** — [10.1016/j.xgen.2024.100655](https://doi.org/10.1016/j.xgen.2024.100655)
  - Type: analogous | Citations: 19
  - Shows the use of knowledge graph embeddings to deconvolute complex biological interactions, which is analogous to using a chemical ontology to disentangle spectral mixtures.
- **[2025] Unifying multi-sample network inference from prior knowledge and omics data with CORNETO** — [10.1038/s42256-025-01069-9](https://doi.org/10.1038/s42256-025-01069-9)
  - Type: analogous | Citations: 16
  - Demonstrates a framework that integrates prior knowledge with data for network inference, supporting the idea of knowledge-augmented models in a different domain.
- **[2008] Gene Ontology term overlap as a measure of gene functional similarity** — [10.1186/1471-2105-9-327](https://doi.org/10.1186/1471-2105-9-327)
  - Type: analogous | Citations: 204
  - Provides a method for measuring semantic similarity using ontology terms, which could be adapted to measure similarity between chemical interferents in a wastewater ontology.
- **[2019] A Neural Named Entity Recognition and Multi-Type Normalization Tool for Biomedical Text Mining** — [10.1109/ACCESS.2019.2920708](https://doi.org/10.1109/ACCESS.2019.2920708)
  - Type: analogous | Citations: 121
  - Demonstrates the use of context and ontologies for entity normalization in text, which is the methodological basis for the proposed semantic normalization in spectral analysis.
- **[2017] FamPlex: a resource for entity recognition and relationship resolution of human protein families and complexes in biomedical text mining** — [10.1186/s12859-018-2211-5](https://doi.org/10.1186/s12859-018-2211-5)
  - Type: analogous | Citations: 34
  - Shows the importance of structured resources for entity grounding in text mining, analogous to the need for a chemical ontology for spectral interpretation.
- **[2024] HunFlair2 in a cross-corpus evaluation of biomedical named entity recognition and normalization tools** — [10.1093/bioinformatics/btae564](https://doi.org/10.1093/bioinformatics/btae564)
  - Type: analogous | Citations: 24
  - Highlights the challenges and importance of robust entity normalization in biomedical text mining, which is directly analogous to the challenge of normalizing spectral signatures.

### 2.3 Contre-preuves et limitations connues

- **[addressable] Extraction of chromium (VI) in water samples by dispersive liquid–liquid microextraction based on deep eutectic solvent and determination by UV–Vis spectrophotometry**
  - This paper uses a pre-concentration/extraction step (DES-DLLME) to isolate Cr(VI) before UV-Vis measurement, suggesting that sample preparation is a viable alternative to computational disentanglement of overlapping spectra.
- **[minor] Functionalization of graphene using deep eutectic solvents**
  - This paper is about graphene functionalization and is not directly related to the hypothesis. Its inclusion in the list is tangential, and it does not provide evidence for or against the hypothesis.
- **[minor] Green Synthesis of a Carbon Quantum Dots-Based Superhydrophobic Membrane for Efficient Oil/Water Separation**
  - This paper focuses on physical separation of oil and water, not on spectral analysis or chemical interferents. It is not relevant to the hypothesis.

### 2.4 Evaluation de nouveaute

- **Score**: 0.8/1.0
- **Verdict**: novel

## 3. Predictions falsifiables

| # | Prediction | Borne quantitative | Methode | H0 | Test statistique |
|---|-----------|-------------------|---------|-----|-----------------|
| 1 | In a controlled ternary mixture (Cu, Zn, Ni), the ontology-augmented model will achieve a lower MAPE than the tuned baseline model. | The ontology-augmented model will achieve a MAPE of ≤ 15% for each metal, which is at least a 20% relative reduction compared to the baseline model's MAPE (e.g., baseline MAPE = 20%, augmented MAPE = 16%). | UV-Vis spectrophotometry (range 200-800 nm) of 100 synthetic samples with known concentrations, analyzed by both models. | H0: The mean MAPE of the ontology-augmented model is equal to or greater than the mean MAPE of the baseline model (μ_augmented ≥ μ_baseline). | One-tailed paired t-test on the per-sample MAPE values, alpha = 0.05, with a pre-registered power analysis (n=100, power=0.80) to detect a 20% relative difference. |
| 2 | When tested on spectra from a different industrial site (external validation), the ontology-augmented model will maintain a MAPE improvement over the baseline model trained on the same external data. | The relative MAPE improvement of the ontology-augmented model over the baseline will be at least 10% on the external site's test set (n=300). | UV-Vis spectra collected from a secondary wastewater treatment plant (Site B) with known concentrations of Cu, Zn, and Ni, analyzed by both models. | H0: The relative MAPE improvement of the ontology-augmented model over the baseline on the external site is ≤ 0%. | Bootstrap confidence interval (95%) for the median relative MAPE improvement. The null is rejected if the lower bound of the CI is > 0%. |
| 3 | When a novel interferent (e.g., Co) is introduced into the test set, the ontology-augmented model will flag a high prediction uncertainty or a 'novelty' score for those samples, while the baseline model will not. | The ontology-augmented model will have a recall of ≥ 0.80 for detecting samples containing the novel interferent, defined as samples where the model's prediction uncertainty (e.g., Monte Carlo dropout variance) exceeds the 95th percentile of the training set's uncertainty distribution. | Introduce Co into 50 of 150 test samples. Measure the model's uncertainty output for each sample. | H0: The recall of the ontology-augmented model for novel interferent detection is ≤ 0.50 (i.e., not better than chance). | One-proportion z-test, alpha = 0.05, comparing observed recall to 0.50. |

## 4. Protocole experimental

**Timeline globale**: 10-18 months
**Budget global**: €25k-120k

### 4.1 Phase 1 — In Silico Validation

**Objectif**: To kill the hypothesis cheaply by testing whether the ontology-augmented model can outperform a tuned baseline on synthetic spectral data, and to validate the core mechanism of semantic disambiguation before any wet-lab work.

**Methodologie**: 1. **Synthetic Data Generation (Python, NumPy/SciPy):** Generate 10,000 UV-Vis spectra (200-800 nm, 1 nm resolution) using the Beer-Lambert law with known molar absorptivity curves for Cu, Zn, Ni, and interferents (e.g., Co, Fe). Add Gaussian noise (SNR 20-50 dB) and random baseline drift. Concentrations: 0.1-10 mg/L. pH: fixed at 5.0 (simulate buffer). Use a random split (70/15/15 train/val/test).
2. **Ontology Construction (Protégé, OWL):** Build a minimal ontology (Phase 0) encoding: (a) spectral peak positions and molar absorptivities for each metal, (b) pH-dependent speciation rules, (c) co-occurrence probabilities derived from a simulated industrial discharge dataset (e.g., Cu and Zn co-occur 80% of the time, Ni rarely with Cu). Export as RDF/OWL.
3. **Baseline Model (scikit-learn):** Train a tuned Elastic Net (via GridSearchCV) on the raw spectra (flattened) to predict [Cu], [Zn], [Ni]. Record MAPE per metal on the test set.
4. **Ontology-Augmented Model (PyTorch Geometric):** Implement a GNN where each spectrum is converted to a graph: nodes = detected peaks (via scipy.signal.find_peaks), node features = peak position, height, width; edges = ontology-derived relationships (e.g., 'overlaps_with', 'co-occurs_with') with edge weights = co-occurrence probability. Train a GraphSAGE or GAT model to predict concentrations. Use Monte Carlo dropout for uncertainty estimation.
5. **Novel Interferent Simulation:** In the test set, introduce 50 samples with Co (not in ontology). Measure the GNN's uncertainty (variance of MC dropout predictions) and the baseline's prediction error.
6. **Statistical Analysis (SciPy/statsmodels):** Perform one-tailed paired t-test on per-sample MAPE (H0: μ_augmented ≥ μ_baseline). Compute recall of novel interferent detection (uncertainty > 95th percentile of training set).

- Cout: €1,500-3,000 (mostly personnel time)
- Duree: 4-8 weeks
- Equipement: High-performance workstation (CPU, 32GB RAM, NVIDIA GPU 8GB+)
- Logiciels: Python 3.10+, PyTorch, PyTorch Geometric, scikit-learn, NumPy/SciPy, Protégé (OWL editor), RDKit (for chemical structure handling), Jupyter Notebooks

**Criteres de succes :**

- MAPE reduction (relative): Ontology-augmented model shows ≥20% relative reduction in MAPE vs. baseline for at least 2 of 3 metals (e.g., baseline 20% → augmented ≤16%)
- Statistical significance: p-value < 0.05 from one-tailed paired t-test (μ_augmented < μ_baseline)
- Novel interferent recall: Recall ≥ 0.80 for detecting Co-containing samples (uncertainty > 95th percentile)

- **GO**: All three success criteria are met (MAPE reduction, significance, and novel interferent recall).
- **NO-GO**: The augmented model does not outperform baseline (p ≥ 0.05) OR recall < 0.60. This suggests the ontology priors are not providing useful inductive bias.
- **PIVOT**: If the baseline is already achieving MAPE < 10%, the test may be underpowered. Consider increasing spectral noise or adding more complex interferent mixtures. If the GNN fails to train, simplify the graph construction (e.g., use fully connected graph with attention).

### 4.2 Phase 2 — Minimal Experimental Validation

**Objectif**: To validate the core mechanism on real, controlled laboratory samples: does the ontology-augmented model outperform the baseline on a small set of synthetic mixtures measured by a real spectrophotometer?

**Methodologie**: 1. **Sample Preparation (Wet Lab):** Prepare 60 synthetic samples in 0.1M acetate buffer (pH 5.0) with known concentrations of Cu, Zn, Ni (0.1-10 mg/L) using standard metal salt solutions (CuSO4, ZnSO4, NiSO4). Include 15 samples with Co (novel interferent) at 0.5-5 mg/L. Use a random design (e.g., Latin hypercube) for concentration ranges.
2. **Spectra Acquisition:** Measure UV-Vis spectra (200-800 nm) using a benchtop spectrophotometer (e.g., Shimadzu UV-1900, Agilent Cary 60) with a 1 cm quartz cuvette. Record 3 technical replicates per sample. Ensure temperature control (25°C).
3. **Data Preprocessing:** Average replicates, subtract buffer blank, apply scatter correction (e.g., standard normal variate, SNV).
4. **Model Adaptation:** Fine-tune the pre-trained models from Phase 1 on this small real dataset (transfer learning). Use 40 samples for training, 20 for testing (including all 15 Co samples in the test set).
5. **Evaluation:** Compare MAPE of baseline vs. augmented model on the 20 test samples. Measure novel interferent detection recall using the GNN's uncertainty.
6. **Statistical Analysis:** Given the small sample size (n=20), use a bootstrap confidence interval for the median MAPE difference. For novel interferent detection, use a one-proportion z-test (H0: recall ≤ 0.50).

- Cout: €4,000-8,000 (consumables, spectrophotometer access, personnel)
- Duree: 1-2 months
- Equipement: UV-Vis spectrophotometer (dual-beam, 200-800 nm), Analytical balance, pH meter, Volumetric flasks and pipettes, Quartz cuvettes (1 cm path length), Standard metal solutions (Cu, Zn, Ni, Co)
- Logiciels: Python (same as Phase 1), Spectrophotometer software (e.g., UVProbe), Excel or similar for sample tracking

**Criteres de succes :**

- MAPE improvement (relative): ≥10% relative improvement of augmented model over baseline on the real test set (e.g., baseline 25% → augmented ≤22.5%)
- Novel interferent recall: Recall ≥ 0.70 for Co detection (uncertainty > 95th percentile of training set)
- Absolute MAPE: Augmented model achieves MAPE ≤ 20% for each metal (Cu, Zn, Ni)

- **GO**: At least two of the three success criteria are met, including the MAPE improvement criterion. This confirms the mechanism works on real data.
- **NO-GO**: No improvement over baseline (CI includes 0 or negative) OR recall < 0.50. This suggests the ontology priors are not transferable to real spectral noise and matrix effects.
- **PIVOT**: If the baseline performs poorly (MAPE > 30%), the issue may be spectral preprocessing, not the ontology. Revisit preprocessing steps. If the augmented model overfits the small dataset, increase regularization or use a simpler semantic layer.

### 4.3 Phase 3 — Full Experimental Protocol

**Objectif**: To rigorously validate the hypothesis under realistic conditions, including external site validation and novel interferent detection, producing publishable results.

**Methodologie**: 1. **Ontology Refinement (Phase 0 extended):** Expand the ontology to include more interferents (Fe, Co, Mn, Cr) and site-specific co-occurrence data from an actual industrial partner (e.g., electroplating facility). Encode pH-dependent speciation for all metals.
2. **Controlled Lab Experiments (Site A):** Generate 300 synthetic samples in the lab (as in Phase 2) with a wider concentration range (0.1-10 mg/L) and more complex mixtures (up to 5 metals). Include 100 samples with novel interferents (e.g., Co, Cr) for detection testing. Measure UV-Vis spectra.
3. **External Validation (Site B):** Collect 300 real wastewater samples from a secondary treatment plant (Site B) with known concentrations of Cu, Zn, Ni (via ICP-MS as reference). Measure UV-Vis spectra on-site or in the lab. This tests the transferability of the ontology built from Site A's co-occurrence data.
4. **Model Training and Evaluation:** Train the baseline and augmented models on Site A data (n=300). Evaluate on: (a) held-out Site A test set (n=100), (b) Site B external set (n=300). For novel interferent detection, use the Site A test set with novel metals.
5. **Statistical Analysis:** 
   - **Prediction 1 (Lab):** One-tailed paired t-test on per-sample MAPE (n=100, alpha=0.05, power=0.80). Pre-register the analysis plan.
   - **Prediction 2 (External):** Bootstrap 95% CI for the median relative MAPE improvement on Site B. Reject H0 if lower bound > 0%.
   - **Prediction 3 (Novel):** One-proportion z-test on recall (H0: recall ≤ 0.50).
6. **Ablation Studies:** To strengthen the causal claim, run the augmented model with (a) full ontology, (b) partial ontology (80% of interferents), (c) no ontology (baseline). This tests the effect of ontology coverage.
7. **Publication:** Prepare a manuscript with full methods, code, and data availability statement.

- Cout: €25,000-80,000 (equipment access, consumables, ICP-MS analysis, personnel, publication fees)
- Duree: 6-12 months
- Equipement: UV-Vis spectrophotometer (high-precision, e.g., Agilent Cary 7000), ICP-MS (for reference concentrations at Site B), pH meter, balance, glassware, Automated liquid handling system (optional, for high-throughput sample prep)
- Logiciels: Python (PyTorch, PyTorch Geometric), R or Python for statistical analysis, Protégé for ontology management, Git for version control, DVC for data versioning

**Criteres de succes :**

- Prediction 1 (Lab MAPE): Augmented model achieves MAPE ≤ 15% for each metal AND ≥20% relative reduction vs. baseline (p < 0.05)
- Prediction 2 (External MAPE): Relative MAPE improvement ≥ 10% on Site B (lower bound of 95% bootstrap CI > 0%)
- Prediction 3 (Novel detection): Recall ≥ 0.80 for novel interferent detection (p < 0.05 vs. 0.50)
- Ablation (Ontology coverage): Full ontology > Partial ontology > Baseline in terms of MAPE (monotonic improvement)

- **GO**: All three primary predictions are confirmed (p < 0.05 or CI excludes null) AND the ablation study shows a monotonic trend. This provides strong evidence for the hypothesis.
- **NO-GO**: Any of the three primary predictions fails. This indicates the ontology augmentation is not robustly beneficial across conditions.
- **PIVOT**: If Prediction 2 fails (external validation), the ontology's co-occurrence priors may be too site-specific. Pivot to a 'universal' ontology with only chemical properties (spectra, pKa) and no site-specific co-occurrence. If Prediction 3 fails, the uncertainty estimation method needs improvement (e.g., use deep ensembles instead of MC dropout).

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

- **Peut demarrer maintenant**: Oui
- **Premiere action**: Generate 1,000 synthetic UV-Vis spectra using known molar absorptivity data for Cu, Zn, Ni (from literature or NIST) and a simple additive model with Gaussian noise. Save as a NumPy array. Simultaneously, download and install PyTorch Geometric and Protégé.
- **Outils**: Python 3.10+, NumPy, SciPy, PyTorch, PyTorch Geometric, scikit-learn, Protégé (OWL editor)
- **Donnees ouvertes**: NIST Chemistry WebBook (for molar absorptivity data of metal complexes), UCI Machine Learning Repository (for general spectral datasets, if any), Zenodo (search for 'UV-Vis wastewater spectra' for any open datasets), PubChem (for chemical structures and properties of metal ions and ligands)

## 5. Analyse d'impact

### 5.1 Impact scientifique

Novelty score: 0.8/1.0 (novel)

### 5.2 Applications industrielles et marche

**Score industriel**: 6.8/10

**Forces :**
- Marche cible identifiable et reglemente : les stations de traitement des eaux usees industrielles et municipales (marche global de l'analyse UV-Vis en ligne estime a ~$1.2B d'ici 2027, CAGR 7-8%) doivent se conformer a des normes de rejet de plus en plus strictes (Directive Cadre sur l'Eau UE, Clean Water Act US). Les exploitants paient deja pour des analyseurs multi-parametres (ex: spectrometres s::can, Hach) et des contrats de maintenance/calibration. Notre proposition de valeur est un logiciel incrementiel qui augmente la precision de leurs actifs existants sans nouvel investissement materiel, ce qui est un argument de vente fort.
- Avantage competitif defensible par la donnee : les ontologies chimiques et les priors de co-occurrence d'interferents sont des actifs de connaissance qui s'accumulent avec chaque site deployee. Un concurrent ne peut pas copier rapidement cette base de connaissance semantique sans acces a des annees de donnees terrain diversifiees. Cela cree une barriere a l'entree par la donnee et le savoir-faire, similaire au modele de Palantir pour l'integration de donnees complexes, mais applique a la chimie analytique.

**Faiblesses :**
- Barriere a l'adoption : le marche cible (exploitants de stations d'epuration) est conservateur et base ses decisions sur des normes et des methodes certifiees (ex: ICP-MS, photometrie). Un modele d'IA 'boite noire' ou meme un modele avec une ontologie complexe devra passer par un processus de validation long aupres des organismes de normalisation (ex: US EPA, AFNOR) pour etre accepte comme methode alternative fiable. Sans certification, les acheteurs potentiels ne remplaceront pas leurs methodes de reference.
- Risque de fragmentation du marche : le protocole se concentre sur Cu, Zn, Ni, Co. Or, les besoins industriels sont beaucoup plus larges (composes organiques, nitrates, phosphates, sulfures). Si l'ontologie doit etre etendue a chaque nouvelle molecule, le cout de maintenance de la base de connaissance pourrait devenir prohibitif et le ROI pour un client specifique trop faible. La commercialisation devra cibler un nombre restreint de 'cas d'usage killer' (ex: metaux lourds dans les effluents miniers) pour etre rentable.

**Recommandation**: Ne pas lancer un produit a grande echelle. Cibler un marche de niche a forte valeur : les industries minieres et metallurgiques (effluents acides charges en metaux) ou les couts de non-conformite sont eleves. Developper un 'module d'expertise' en tant que fonctionnalite premium ajoutee a un logiciel d'analyse spectrale existant. La strategie est de signer 2-3 contrats pilotes avec des acteurs industriels (ex: grands groupes miniers) des la fin de la Phase 2, en vendant une promesse de reduction des couts d'analyse labo (ICP-MS) de 30%. Le ROI pour le client est un retour sur investissement en moins de 18 mois via la reduction des analyses de laboratoire externes et l'optimisation des processus de traitement. Ne pas viser le marche municipal (trop reglemente et conservateur) dans un premier temps.

### 5.3 Opportunites de financement

| Programme | Agence | Fit | Budget type | Taux succes |
|-----------|--------|-----|-------------|-------------|
| ANR Jeune Chercheuse - Jeune Chercheur (JCJC) 2026 | Agence Nationale de la Recherche (France) | 0.85 | 150k€ - 300k€ (pour 36-48 mois) | ~15-20% (taux de succes ANR generique 2024) |
| ERC Starting Grant 2026 (StG) | European Research Council | 0.6 | 1.5M€ - 2.5M€ (sur 5 ans) | ~14% (taux de succes StG 2024) |
| EIC Pathfinder Open 2025 (Horizon Europe) | European Innovation Council (Commission Europeenne) | 0.45 | 3M€ - 5M€ (consortium de 3-5 partenaires) | ~5-8% (taux de succes Pathfinder Open 2024) |

- **ANR Jeune Chercheuse - Jeune Chercheur (JCJC) 2026** (Agence Nationale de la Recherche (France)): Le format JCJC est concu pour des projets a risque et a prise de risque scientifique. Le budget demande (120k€) est inferieur au plafond, ce qui est un atout (projet frugal). Le protocole en 3 phases avec GO/NO-GO correspond exactement a l'esprit 'risque controle' de l'ANR. L'absence de besoin de consortium lourd est un avantage. Le seul risque est la concurrence sur le sujet 'IA pour l'environnement' qui est tres competitif.
- **ERC Starting Grant 2026 (StG)** (European Research Council): L'ERC StG est le seul programme capable de financer une exploration fondamentale de longue haleine. Le projet est un excellent candidat pour le pilier 'Physical Sciences & Engineering' (PE6) ou 'Information Science' (PE7). Cependant, le budget demande est 10 fois inferieur au plafond, ce qui est un signal negatif (manque d'ambition). Il faudrait redimensionner le projet pour inclure une extension vers d'autres types de spectroscopie (NIR, Raman) ou d'autres matrices environnementales (sol, air) pour justifier un budget plus important. Le TRL initial est trop faible pour l'ERC, qui finance des idees, mais il faut un narrative plus large.
- **EIC Pathfinder Open 2025 (Horizon Europe)** (European Innovation Council (Commission Europeenne)): Ce programme finance des technologies radicalement nouvelles a TRL 2-4, mais il exige un consortium multi-partenaires (academie + industrie). Le projet pourrait etre redimensionne pour inclure un partenaire industriel (ex: Endress+Hauser, Veolia) et un expert en ontologies (ex: laboratoire d'informatique). Le budget de 120k€ est totalement inadapte, il faudrait le multiplier par 10. Le programme valorise les 'deep tech' a fort impact societale (eau propre), ce qui colle, mais la maturite du consortium est un point bloquant. A considerer uniquement si vous trouvez un partenaire industriel solide.

**Recommandation financement**: Ciblez en priorite un financement exploratoire et individuel : l'ANR Jeune Chercheuse (JCJC) est le meilleur fit pour demarrer la Phase 1 et 2 (budget 120k€ est parfait). Parallelement, preparez un dossier ERC Starting Grant en positionnant l'hypothese comme une rupture methodologique en 'AI for Science' (attention, le budget est trop faible pour un ERC, il faudra le justifier comme un projet pilote). Evitez Horizon Europe Pillar 2 a ce stade, car le TRL est trop bas et le consortium trop restreint. Utilisez la Phase 1 (in silico) comme preuve de concept pour renforcer le dossier ERC/JCJC.

## 6. Panel Review Summary

| Reviewer | Score | Verdict | Point cle |
|----------|-------|---------|-----------|
| methodologist | 7.8/10 | accept | The progressive validation strategy (in silico → minimal wet lab → full external validation) is exemplary. It explicitly builds in a 'kill criterion' in Phase 1 (cheaply testing the core mechanism before committing significant wet-lab resources), which is a hallmark of rigorous experimental design and directly addresses the risk of wasting resources on a fundamentally flawed hypothesis. |
| domain_expert | 7.8/10 | accept | The hypothesis is theoretically coherent and aligns strongly with the emerging KA-ML paradigm. The proposition that ontology acts as an inductive bias to constrain the hypothesis space is a well-established concept in statistical learning theory, and its application to spectral deconvolution is a logical and timely extension. |
| contrarian | 3.5/10 | weak_reject | The explicit acknowledgment of the external validation problem (Site B) and the novelty detection problem (Co interferent) is commendable. Most spectral ML papers ignore these practical deployment issues entirely, and the authors at least attempt to define testable bounds for them. |
| industrialist | 6.8/10 | accept | Marche cible identifiable et reglemente : les stations de traitement des eaux usees industrielles et municipales (marche global de l'analyse UV-Vis en ligne estime a ~$1.2B d'ici 2027, CAGR 7-8%) doivent se conformer a des normes de rejet de plus en plus strictes (Directive Cadre sur l'Eau UE, Clean Water Act US). Les exploitants paient deja pour des analyseurs multi-parametres (ex: spectrometres s::can, Hach) et des contrats de maintenance/calibration. Notre proposition de valeur est un logiciel incrementiel qui augmente la precision de leurs actifs existants sans nouvel investissement materiel, ce qui est un argument de vente fort. |
| funding_strategist | 6.5/10 | weak_accept | Originalite conceptuelle forte : l'application d'ontologies chimiques structurees comme contrainte inductive pour la deconvolution spectrale UV-Vis est une approche peu exploree, se distinguant des methodes de deep learning classiques (CNN, autoencoders) qui dominent le domaine. Ce positionnement 'knowledge-augmented ML' est tres attractif pour les evaluateurs ERC/ANR cherchant des ruptures methodologiques. |

**Consensus score**: 6.5/10
**Verdict final**: publish_brief

### 6.1 Consensus

- La stratégie de validation progressive en 3 phases avec critères GO/NO-GO explicites est unanimement saluée comme un modèle de rigueur expérimentale, minimisant les risques financiers et scientifiques.
- L'hypothèse est bien formulée, falsifiable et théoriquement ancrée dans le paradigme émergent du knowledge-augmented ML, avec un positionnement clair par rapport à l'état de l'art.
- Le principal risque identifié par tous (à l'exception du contrarian qui le juge rédhibitoire) est la transférabilité des priors de co-occurrence entre sites industriels, qui conditionne la validité de la validation externe.

### 6.2 Points de desaccord

- Le contrarian juge l'approche comme un 'marteau-pilon pour écraser une noix' (sur-ingénierie pour un problème 3-composants soluble par chimiométrie classique), tandis que les autres reviewers y voient une avancée conceptuelle légitime et généralisable.
- Désaccord sur la pertinence du baseline : le contrarian exige une comparaison avec des modèles bayésiens (GPR, BNN) pour la détection de nouveauté, alors que le methodologist se concentre sur la spécification rigoureuse du tuning de l'Elastic Net.
- Le funding strategist et l'industrialist divergent sur la cible : le premier recommande un financement académique exploratoire (ANR JCJC), le second une commercialisation de niche immédiate vers l'industrie minière, reflétant une tension entre ambition scientifique et viabilité commerciale.

### 6.3 Critical path

La démonstration que l'ontologie chimique apporte un gain prédictif réel et non trivial par rapport à un baseline fort (et non sous-ajusté) sur des données synthétiques, ET que les priors de co-occurrence peuvent être dynamiquement adaptés ou désactivés face à un site externe aux caractéristiques différentes, sans dégradation catastrophique des performances.

**Recommandation finale**: Le panel recommande la publication du brief malgré les réserves substantielles du contrarian. La rigueur méthodologique du protocole, la clarté de l'hypothèse et le potentiel de rupture conceptuelle justifient de poursuivre l'investigation. Les inquiétudes du contrarian sur la sur-ingénierie et le risque de transfert négatif sont sérieuses mais ne sont pas rédhibitoires à ce stade ; elles doivent être intégrées comme des risques critiques à tester explicitement dès la Phase 1 (in silico). Le brief doit être publié tel quel, avec la recommandation forte que les expériences initiales comparent systématiquement le modèle ontologique à des baselines bayésiennes et testent la robustesse des priors face à des distributions de co-occurrence dégradées.

## 7. Gap Manifest residuel

### 7.1 Data gaps

- No existing chemical ontology for wastewater pollutants with spectral properties and co-occurrence rules.

### 7.2 Competence gaps

- Phase 1: Python programming
- Phase 1: Machine learning (GNNs, regularization)
- Phase 1: Chemometrics (Beer-Lambert law)
- Phase 1: Ontology engineering (OWL/RDF)
- Phase 1: Statistical hypothesis testing
- Phase 2: Wet lab chemistry (solution preparation, pH buffering)
- Phase 2: Spectrophotometry
- Phase 2: Data preprocessing (SNV, baseline correction)
- Phase 2: Transfer learning (fine-tuning neural networks)
- Phase 3: Advanced chemometrics
- Phase 3: Environmental sampling and analysis (wastewater)
- Phase 3: ICP-MS operation
- Phase 3: Machine learning engineering (GNNs, deployment)
- Phase 3: Statistical design of experiments
- Phase 3: Scientific writing

### 7.3 Epistemic gaps

- We do not know if the ontology's co-occurrence priors, built from one industrial site, will be sufficiently generalizable to improve predictions at a different site with different discharge profiles.
- We do not know the minimum size and completeness of the ontology required to achieve a significant improvement over the baseline.
- We do not know if the model can effectively flag a novel interferent (not in the ontology) without a significant drop in accuracy for known analytes.

## References

[1] Fengbo Zhou, Changgeng Li, et al. (2019). *Simultaneous determination of trace metal ions in industrial wastewater based on UV–vis spectrometry*. DOI: [10.1016/J.IJLEO.2018.09.075](https://doi.org/10.1016/J.IJLEO.2018.09.075)
[2] Christel Sirocchi, Alessandro Bogliolo, et al. (2024). *Medical-informed machine learning: integrating prior knowledge into medical decision systems*. DOI: [10.1186/s12911-024-02582-4](https://doi.org/10.1186/s12911-024-02582-4)
[3] Katharina Beckh, Sebastian Müller, et al. (2023). *Harnessing Prior Knowledge for Explainable Machine Learning: An Overview*. DOI: [10.1109/SaTML54575.2023.00038](https://doi.org/10.1109/SaTML54575.2023.00038)
[4] A. Gray, J. Sadler, et al. (2011). *A Semantic Sensor Web for Environmental Decision Support Applications*. DOI: [10.3390/s110908855](https://doi.org/10.3390/s110908855)
[5] Jude A. Adeleke, Deshendran Moodley, et al. (2017). *Integrating Statistical Machine Learning in a Semantic Sensor Web for Proactive Monitoring and Control*. DOI: [10.3390/s17040807](https://doi.org/10.3390/s17040807)
[6] Susel Fernández, Iván Marsa-Maestre, et al. (2013). *Ontology Alignment Architecture for Semantic Sensor Web Integration*. DOI: [10.3390/s130912581](https://doi.org/10.3390/s130912581)
[7] Shengkun Ni, Xiangtai Kong, et al. (2024). *Identifying compound-protein interactions with knowledge graph embedding of perturbation transcriptomics*. DOI: [10.1016/j.xgen.2024.100655](https://doi.org/10.1016/j.xgen.2024.100655)
[8] P. Rodríguez-Mier, Martín Garrido-Rodríguez, et al. (2025). *Unifying multi-sample network inference from prior knowledge and omics data with CORNETO*. DOI: [10.1038/s42256-025-01069-9](https://doi.org/10.1038/s42256-025-01069-9)
[9] Meeta Mistry, P. Pavlidis (2008). *Gene Ontology term overlap as a measure of gene functional similarity*. DOI: [10.1186/1471-2105-9-327](https://doi.org/10.1186/1471-2105-9-327)
[10] Donghyeon Kim, Jinhyuk Lee, et al. (2019). *A Neural Named Entity Recognition and Multi-Type Normalization Tool for Biomedical Text Mining*. DOI: [10.1109/ACCESS.2019.2920708](https://doi.org/10.1109/ACCESS.2019.2920708)
[11] John A. Bachman, B. Gyori, et al. (2017). *FamPlex: a resource for entity recognition and relationship resolution of human protein families and complexes in biomedical text mining*. DOI: [10.1186/s12859-018-2211-5](https://doi.org/10.1186/s12859-018-2211-5)
[12] Mario Sänger, Samuele Garda, et al. (2024). *HunFlair2 in a cross-corpus evaluation of biomedical named entity recognition and normalization tools*. DOI: [10.1093/bioinformatics/btae564](https://doi.org/10.1093/bioinformatics/btae564)
[13] Mahsa Pourmohammad, M. Faraji, et al. (?). *Extraction of chromium (VI) in water samples by dispersive liquid–liquid microextraction based on deep eutectic solvent and determination by UV–Vis spectrophotometry*. DOI: [10.1080/03067319.2019.1650920](https://doi.org/10.1080/03067319.2019.1650920)
[14] M. Hayyan, Ali Abo-Hamad, et al. (2015). *Functionalization of graphene using deep eutectic solvents*. DOI: [10.1186/s11671-015-1004-2](https://doi.org/10.1186/s11671-015-1004-2)
[15] R. Almufarij, M. E. Mohamed (2023). *Green Synthesis of a Carbon Quantum Dots-Based Superhydrophobic Membrane for Efficient Oil/Water Separation*. DOI: [10.3390/ma16155456](https://doi.org/10.3390/ma16155456)

## Annexes

### A. Detailed Reviewer Reports

#### Methodologist

- **Score**: 7.8/10 | **Verdict**: accept | **Confidence**: 0.88
- **Strengths**: The progressive validation strategy (in silico → minimal wet lab → full external validation) is exemplary. It explicitly builds in a 'kill criterion' in Phase 1 (cheaply testing the core mechanism before committing significant wet-lab resources), which is a hallmark of rigorous experimental design and directly addresses the risk of wasting resources on a fundamentally flawed hypothesis.; The hypothesis is exceptionally well-formulated and falsifiable. The specification of quantitative bounds (e.g., ≥20% relative MAPE reduction, recall ≥0.80) and explicit null hypotheses for each prediction transforms a vague claim into a testable statistical framework. This pre-registration of effect sizes is a major strength that directly combats confirmation bias.; The inclusion of an external validation phase (Site B) and an ablation study (full vs. partial ontology) is a sophisticated and robust approach. The external validation tests generalizability and guards against overfitting to site-specific noise, while the ablation study is critical for establishing a causal link between ontology coverage and performance, rather than just a correlation.
- **Weaknesses**: The protocol for the baseline model is under-specified. While it mentions a 'tuned Elastic Net via GridSearchCV', it does not specify the cross-validation strategy (e.g., k-fold, repeated) or the hyperparameter grid. This lack of detail threatens the fairness of the comparison; a poorly tuned baseline could inflate the apparent benefit of the augmented model, undermining the validity of the primary comparison.; The statistical analysis for the primary endpoint (MAPE comparison) relies on a paired t-test, which assumes a normal distribution of the per-sample MAPE differences. Spectral prediction errors are often heteroscedastic and non-normal, especially with varying concentrations. The protocol should pre-specify a robust alternative (e.g., Wilcoxon signed-rank test) or a transformation plan, to ensure the validity of the p-value.; The protocol does not adequately address the potential for measurement bias in the external validation (Site B). It states that concentrations are 'known via ICP-MS', but it does not detail how the ICP-MS data will be handled (e.g., values below detection limit, measurement error). If the reference method has significant error, it will add noise to the evaluation, potentially biasing the results toward the null and making the test underpowered.
- **Questions**: For the baseline model (Elastic Net), what is the exact cross-validation scheme (e.g., 5-fold, 10-fold, repeated) and the specific hyperparameter grid (e.g., range of alpha and l1_ratio) that will be used for tuning? How will you ensure that the baseline is not under-tuned, which would unfairly favor the augmented model?; Given that spectral errors are often non-normal, what is the pre-specified plan for the statistical analysis of the primary endpoint if the assumptions of the paired t-test are violated? Will you use a non-parametric alternative or a data transformation, and what are the criteria for making that decision?; In the external validation (Phase 3), how will you handle ICP-MS measurements that are below the limit of detection (LOD) or quantification (LOQ)? Will these samples be excluded, imputed, or treated as censored data, and how might this impact the comparison of MAPE between the two models?
- **Recommendation**: Accept with minor revisions. The protocol is exceptionally well-designed with a clear, falsifiable hypothesis and a strong progressive validation structure. The primary concerns are the lack of detail on the baseline model tuning and the potential violation of statistical assumptions. I strongly recommend the authors pre-specify the exact tuning procedure, the statistical analysis plan for non-normal data, and the handling of ICP-MS censored data before proceeding with Phase 1.

#### Domain Expert

- **Score**: 7.8/10 | **Verdict**: accept | **Confidence**: 0.85
- **Strengths**: The hypothesis is theoretically coherent and aligns strongly with the emerging KA-ML paradigm. The proposition that ontology acts as an inductive bias to constrain the hypothesis space is a well-established concept in statistical learning theory, and its application to spectral deconvolution is a logical and timely extension.; The mechanistic pipeline is plausible and well-articulated. The progression from ontology construction (Phase 0) to graph encoding and GNN-based processing is a technically sound architecture. The use of co-occurrence priors as a regularizer to penalize chemically implausible concentration combinations is a novel and chemically meaningful mechanism, directly addressing the ill-posed nature of spectral unmixing.; The positioning against the literature is correct. The work correctly identifies the gap between classical signal processing (e.g., wavelet transform for metal ion overlap) and modern knowledge-augmented ML. By citing analogous work in medical informatics and knowledge graph embeddings for biological deconvolution, it establishes a strong precedent for the proposed approach, correctly positioning the novelty as the transfer of this paradigm to chemical spectral analysis.
- **Weaknesses**: The critical assumption of transferability of co-occurrence priors across different industrial sites is a major risk. The hypothesis acknowledges this as a 'known unknown', but it is a fundamental limitation. If co-occurrence patterns are site-specific, the ontology may introduce a strong, incorrect bias that could degrade performance on new data, undermining the core claim of improved generalization.; The mechanism for handling novel interferents is under-specified. The hypothesis states it is a 'known unknown' whether the model can flag a novel interferent, but the proposed architecture (GNN over a fixed ontology graph) seems inherently incapable of this. A GNN will likely force a novel spectral feature into an existing node or treat it as noise, potentially leading to significant errors in predicting known analytes. This is a critical failure mode that needs a more concrete mitigation strategy.; The 'semantic normalization' mechanism is not clearly defined. While the steps describe graph encoding and attention, the precise mathematical or algorithmic form of the 'normalization' is vague. It is unclear how the ontology-derived weights are exactly integrated into the loss function or the attention mechanism to provide a 'chemically plausible' constraint, leaving a gap between the conceptual idea and its implementable form.
- **Questions**: Given the acknowledged site-specificity of co-occurrence priors, what is the proposed methodology for calibrating or adapting the ontology's relational weights when deploying the model to a new site with a different discharge profile? Is a fine-tuning step envisioned, or is the ontology considered static?; How does the GNN architecture explicitly differentiate between a spectral feature that is a known interferent with low co-occurrence probability and a completely novel interferent? What specific output or internal state would flag this novelty, and how would the model be trained to recognize the difference between a 'known-absent' and an 'unknown' spectral signature?; Could you provide a more formal definition of the 'semantic normalization' regularizer? Specifically, how is the penalty for a chemically implausible combination (e.g., high Cu and high Ni) formulated in the loss function, and how are the co-occurrence probabilities from the ontology converted into a differentiable constraint?
- **Recommendation**: The hypothesis presents a compelling and novel direction for UV-Vis spectral analysis, effectively bridging the gap between chemical ontologies and modern graph-based ML. The theoretical framing is strong, and the proposed mechanism is a significant step beyond current signal-processing methods. However, the success hinges on resolving the critical issue of prior transferability and defining a robust mechanism for handling novel interferents. I recommend accepting the hypothesis for further investigation, with a strong emphasis on designing experiments that explicitly test the robustness of the ontology's co-occurrence priors and the model's out-of-ontology detection capabilities.

#### Contrarian

- **Score**: 3.5/10 | **Verdict**: weak_reject | **Confidence**: 0.85
- **Strengths**: The explicit acknowledgment of the external validation problem (Site B) and the novelty detection problem (Co interferent) is commendable. Most spectral ML papers ignore these practical deployment issues entirely, and the authors at least attempt to define testable bounds for them.; The causal chain is clearly articulated, and the assumption of Beer-Lambert additivity is stated upfront, which allows a reviewer to immediately assess the scope of applicability.
- **Weaknesses**: FAIL REASON #1: The 'ontology' is a glorified prior on a 3-component system. For Cu, Zn, Ni in a buffer, the spectral overlap is a known, static problem solvable by classical chemometrics (e.g., PLS, MCR-ALS) with far less complexity. The proposed GNN/ontology framework is massive overkill for a problem where the 'interferent co-occurrence priors' are trivially encoded as a 3x3 correlation matrix. The claimed improvement over a *tuned* Elastic Net is likely to be marginal (<5% MAPE reduction) and well within the noise of spectral preprocessing choices (e.g., scatter correction, wavelength selection). The effect size predicted (20% relative reduction) is unrealistic for such a low-dimensional problem.; FAIL REASON #2: The external validation (Prediction 2) is fundamentally flawed by the 'site-specific co-occurrence' admission. The ontology encodes priors from Site A (e.g., 'high Cu implies low Ni'). If Site B has a different process, these priors become *negative* transfer. The model will be forced to either ignore the ontology (reverting to baseline performance) or actively fight it, leading to *worse* performance than the baseline. The paper does not propose a mechanism for adapting or weighting the ontology's co-occurrence priors based on incoming data, making the 10% improvement bound on Site B a wishful guess, not a testable hypothesis.; FAIL REASON #3: The novelty detection (Prediction 3) is a strawman. The baseline model (LASSO/Elastic Net) is a point-estimate model; it *cannot* flag uncertainty by design. Comparing its lack of uncertainty against a Monte Carlo dropout GNN is a rigged game. The correct baseline for novelty detection is a Gaussian Process Regressor (GPR) or a Bayesian Neural Network (BNN) with calibrated uncertainty. The hypothesis does not compare against these, so the claim that the ontology *specifically* enables novelty detection is unsubstantiated. The ontology provides no mechanism for 'unknown unknowns'; it only knows what is in its graph. A novel interferant (Co) will simply create a new, unmodeled spectral region that the GNN will try to force into existing nodes, likely producing *confident* wrong answers, not high uncertainty.
- **Questions**: If the ontology is built on fundamental chemical properties (molar absorptivity, pKa) that are transferable, why do you need co-occurrence priors from industrial data at all? If you remove the co-occurrence priors and just use a spectral library matching (e.g., via a learned attention over known pure spectra), do you still get the claimed improvement? If not, you are just memorizing site-specific correlations, not 'semantic normalization'.; What is the baseline for the external validation? If the baseline model is trained *only* on Site B data, and the ontology-augmented model is pre-trained on Site A + Site B, the comparison is unfair. You must specify that the ontology-augmented model is *also* retrained on Site B data, but with the ontology as a fixed, untrainable prior. If the ontology cannot be updated, it will be a liability. If it can be updated, what is the update rule, and how does it differ from just re-estimating the LASSO penalty weights?
- **Recommendation**: The authors must first demonstrate a significant improvement over a non-ontology spectral baseline (e.g., a 1D CNN or a GPR with a learned kernel) on the *synthetic* ternary mixture. Then, they must show that the ontology's co-occurrence priors can be *dynamically reweighted* or *discarded* when confronted with a mismatched external site, without catastrophic failure. Finally, they must compare novelty detection against a Bayesian baseline (e.g., GPR) and show that the ontology provides a *specific* advantage in flagging out-of-ontology interferents, not just a generic uncertainty threshold. Without these three demonstrations, the paper is a complex solution to a simple problem with a high risk of overfitting to a single dataset.

#### Industrialist

- **Score**: 6.8/10 | **Verdict**: accept | **Confidence**: 0.72
- **Strengths**: Marche cible identifiable et reglemente : les stations de traitement des eaux usees industrielles et municipales (marche global de l'analyse UV-Vis en ligne estime a ~$1.2B d'ici 2027, CAGR 7-8%) doivent se conformer a des normes de rejet de plus en plus strictes (Directive Cadre sur l'Eau UE, Clean Water Act US). Les exploitants paient deja pour des analyseurs multi-parametres (ex: spectrometres s::can, Hach) et des contrats de maintenance/calibration. Notre proposition de valeur est un logiciel incrementiel qui augmente la precision de leurs actifs existants sans nouvel investissement materiel, ce qui est un argument de vente fort.; Avantage competitif defensible par la donnee : les ontologies chimiques et les priors de co-occurrence d'interferents sont des actifs de connaissance qui s'accumulent avec chaque site deployee. Un concurrent ne peut pas copier rapidement cette base de connaissance semantique sans acces a des annees de donnees terrain diversifiees. Cela cree une barriere a l'entree par la donnee et le savoir-faire, similaire au modele de Palantir pour l'integration de donnees complexes, mais applique a la chimie analytique.
- **Weaknesses**: Barriere a l'adoption : le marche cible (exploitants de stations d'epuration) est conservateur et base ses decisions sur des normes et des methodes certifiees (ex: ICP-MS, photometrie). Un modele d'IA 'boite noire' ou meme un modele avec une ontologie complexe devra passer par un processus de validation long aupres des organismes de normalisation (ex: US EPA, AFNOR) pour etre accepte comme methode alternative fiable. Sans certification, les acheteurs potentiels ne remplaceront pas leurs methodes de reference.; Risque de fragmentation du marche : le protocole se concentre sur Cu, Zn, Ni, Co. Or, les besoins industriels sont beaucoup plus larges (composes organiques, nitrates, phosphates, sulfures). Si l'ontologie doit etre etendue a chaque nouvelle molecule, le cout de maintenance de la base de connaissance pourrait devenir prohibitif et le ROI pour un client specifique trop faible. La commercialisation devra cibler un nombre restreint de 'cas d'usage killer' (ex: metaux lourds dans les effluents miniers) pour etre rentable.
- **Questions**: Quel est le modele de tarification et le cout d'integration reel chez un client pilote ? Un client paiera-t-il un abonnement SaaS annuel (ex: €15k-30k/site) ou une licence perpetuelle ? Quel est le cout d'ingenierie pour connecter votre solution a un spectrometre existant (protocoles de communication, formats de donnees proprietaires) ?; Pouvez-vous garantir une amelioration de precision sur des matrices reelles complexes (boues activees, effluents charges en matieres organiques) qui ne sont pas dans votre protocole de validation ? Le protocole utilise des melanges synthetiques en laboratoire, ce qui est un environnement 'propre'. Un echec sur un effluent reel non filtre tuerait la credibilite commerciale.; Quelle est votre strategie de propriete intellectuelle ? Un brevet sur le 'semantic normalization framework' est-il envisageable, ou est-ce trop algorithmique et donc non brevetable ? Si vous ne pouvez pas proteger l'ontologie elle-meme, comment empecher un grand acteur (ex: Endress+Hauser, Xylem) de copier la fonctionnalite une fois qu'elle est publiee dans un journal scientifique ?
- **Recommendation**: Ne pas lancer un produit a grande echelle. Cibler un marche de niche a forte valeur : les industries minieres et metallurgiques (effluents acides charges en metaux) ou les couts de non-conformite sont eleves. Developper un 'module d'expertise' en tant que fonctionnalite premium ajoutee a un logiciel d'analyse spectrale existant. La strategie est de signer 2-3 contrats pilotes avec des acteurs industriels (ex: grands groupes miniers) des la fin de la Phase 2, en vendant une promesse de reduction des couts d'analyse labo (ICP-MS) de 30%. Le ROI pour le client est un retour sur investissement en moins de 18 mois via la reduction des analyses de laboratoire externes et l'optimisation des processus de traitement. Ne pas viser le marche municipal (trop reglemente et conservateur) dans un premier temps.

#### Funding Strategist

- **Score**: 6.5/10 | **Verdict**: weak_accept | **Confidence**: 0.7
- **Strengths**: Originalite conceptuelle forte : l'application d'ontologies chimiques structurees comme contrainte inductive pour la deconvolution spectrale UV-Vis est une approche peu exploree, se distinguant des methodes de deep learning classiques (CNN, autoencoders) qui dominent le domaine. Ce positionnement 'knowledge-augmented ML' est tres attractif pour les evaluateurs ERC/ANR cherchant des ruptures methodologiques.; Protocole de validation progressif exemplaire : la structure en 3 phases avec des criteres GO/NO-GO explicites et un budget initial minimal (Phase 1 a ~3k€) demontre une rigueur scientifique rare et un excellent rapport cout/risque. Cela correspond exactement a la philosophie des financements exploratoires (type ERC Proof of Concept ou ANR Jeune Chercheuse).; Impact applique concret et mesurable : le ciblage de la surveillance des eaux usees par UV-Vis repond a un besoin reglementaire europeen (Directive Cadre sur l'Eau) et operationnel (stations d'epuration). La promesse d'une meilleure prediction des analytes individuels en presence d'interferents est un argument de valorisation solide.
- **Weaknesses**: TRL initial faible (TRL 2-3) : l'hypothese repose sur des simulations in silico. Pour des programmes comme Horizon Europe (Pillar 2) ou des appels bilateraux (ANR PRC), le manque de preuve experimentale prealable (spectres reels, echantillons de terrain) est un handicap majeur. Les evaluateurs de ces programmes exigent generalement un TRL 4-5 minimum.; Consortium potentiel sous-dimensionne : l'hypothese est portee par un seul chercheur (ou une petite equipe). Elle necessite des competences complementaires en chimie analytique (spectroscopie), en ingenierie des connaissances (ontologies, Web semantique) et en machine learning. L'absence d'un partenaire 'chimiste des procedes' ou 'exploitant de station d'epuration' affaiblit la credibilite de la phase 3 (validation externe).; Risque de 'narrative gap' sur la nouveaute de l'ontologie : la formulation suppose l'existence ou la creation d'une ontologie dediee. Si l'ontologie doit etre construite de zero, le cout et le temps de developpement sont sous-estimes (le budget max de 120k€ est insuffisant pour un travail ontologique rigoureux). Il faut clarifier si l'on reuse une ontologie existante (ex: ChEBI, SSN) ou si l'on en cree une nouvelle.
- **Questions**: Positionnement strategique : visez-vous un financement 'bottom-up' (ERC, ANR JCJC) pour explorer la science fondamentale, ou un financement 'top-down' (Horizon Europe Cluster 6, appels EIC) pour resoudre un probleme industriel ? Votre protocole actuel est trop academique pour Horizon Europe, mais trop applique pour un ERC Starting Grant. Quelle est votre priorite ?; Quelle est la strategie de gestion des donnees et de l'ontologie ? Pour convaincre un evaluateur, il faut specifier comment l'ontologie sera maintenue, versionnee et partagee (ex: depot public, alignement avec des standards comme QUDT ou OntoCAPE). Un plan de gestion des donnees (DMP) solide est un plus pour l'ANR et Horizon Europe.; Quel est l'argument de rupture par rapport aux methodes de 'physics-informed neural networks' (PINNs) ? Les PINNs encodent deja des contraintes physiques (lois de Beer-Lambert). Il faut explicitement demontrer en quoi une ontologie chimique (relations entre molecules, co-occurrence) apporte un avantage differentiel par rapport a ces approches deja populaires.
- **Recommendation**: Ciblez en priorite un financement exploratoire et individuel : l'ANR Jeune Chercheuse (JCJC) est le meilleur fit pour demarrer la Phase 1 et 2 (budget 120k€ est parfait). Parallelement, preparez un dossier ERC Starting Grant en positionnant l'hypothese comme une rupture methodologique en 'AI for Science' (attention, le budget est trop faible pour un ERC, il faudra le justifier comme un projet pilote). Evitez Horizon Europe Pillar 2 a ce stade, car le TRL est trop bas et le consortium trop restreint. Utilisez la Phase 1 (in silico) comme preuve de concept pour renforcer le dossier ERC/JCJC.

### B. Semantic Scholar Search Queries

- [novelty] `ontology semantic normalization spectral overlap` — Direct check if the core hypothesis (ontology-based semantic normalization applied to spectral overlap) already exists
- [novelty] `knowledge graph UV-Vis wastewater disentanglement` — Searches for existing work combining knowledge graphs with UV-Vis spectral mixture analysis in wastewater
- [novelty] `semantic reasoning spectral mixture analysis` — Targets the specific mechanism of using semantic reasoning layers to disambiguate spectral mixtures
- [novelty] `chemical ontology interferent spectral deconvolution` — Checks for prior use of chemical ontologies to resolve interferent contributions in spectral data
- [evidence] `biomedical text mining named entity normalization ambiguity resolution` — Provides foundational evidence for the core mechanism of context-aware disambiguation in biomedical NLP
- [evidence] `knowledge graph embedding chemical compound disambiguation` — Evidence for how structured knowledge graphs resolve chemical entity ambiguity, a key transferable mechanism
- [evidence] `UV-Vis spectroscopy overlapping absorption spectra multivariate calibration` — Evidence for the current state-of-the-art in handling spectral overlap in wastewater monitoring (the problem domain)
- [evidence] `prior knowledge machine learning spectral regression` — Evidence for how prior knowledge (ontological or otherwise) can augment ML models in spectral analysis
- [cross_domain] `transfer learning biomedical NLP environmental monitoring` — Finds precedents for transferring biomedical NLP methods to environmental science applications
- [cross_domain] `ontology-driven machine learning analytical chemistry` — Searches for interdisciplinary work where ontologies guide ML models in chemistry/analytical contexts
- [cross_domain] `knowledge-augmented spectral interpretation water quality` — Targets the intersection of knowledge-based AI and water quality spectral analysis specifically
- [cross_domain] `semantic web technologies environmental sensor data` — Finds precedents for using semantic web/ontology technologies in environmental monitoring data interpretation

---

*Generated by SPORE (Systeme de Production d'Opportunites de Recherche par Exploration) on 2026-08-17*