in silico
Phase 1 : In Silico Validation
- Objectif
- Determine whether the proposed SSL pretext tasks (masked-band + next-spectrum) can be simulated on synthetic and/or open UV-VIS time-series data to produce a measurable ΔR² ≥ 0.15 at n=20 versus from-scratch training, and whether the intrinsic dimension k of SSL embeddings is lower than that of from-scratch embeddings. Kill the hypothesis cheaply if the mechanism does not reproduce in silico.
- Coût estimé
- €500-2000 (GPU cloud credits if no local GPU; open datasets free)
- Durée estimée
- 4-8 weeks
- Critères de réussite
- ΔR² (SSL vs from-scratch) at n=20 on synthetic data · ≥ 0.15 with 95% CI lower bound > 0.05 · (Paired Wilcoxon signed-rank test on 5-fold temporal-block CV R², one-sided, α=0.0125)
- ΔR² (SSL vs random-noise pretext) at n=20 · ≥ 0.10 · (Paired Wilcoxon signed-rank test, one-sided, α=0.0125)
- ΔR² (SSL vs PLS-expert) at n=20 · ≥ 0.10 · (Paired Wilcoxon signed-rank test, one-sided, α=0.0125)
- Intrinsic dimension k (SSL vs scratch) · k_SSL ≤ 0.6 × k_scratch · (PCA 95% variance on Gram matrix, bootstrap 1000 iterations, two-sided, α=0.0125)
- Linear probing R² on A254/A400 · R²_probe between 0.40 and 0.70 of R²_SSL (if ≥0.90, shortcut flag) · (Ridge regression on A254/A400 + temporal derivatives, same folds)
- On continue si
- ΔR² (SSL vs scratch) ≥ 0.15 with 95% CI lower bound > 0.05 AND ΔR² (SSL vs random-noise) ≥ 0.10 AND k_SSL ≤ 0.6 × k_scratch on synthetic/open data
- On arrête si
- ΔR² (SSL vs scratch) < 0.05 OR ΔR² (SSL vs random-noise) < 0.05 OR k_SSL > 0.9 × k_scratch on synthetic/open data
- On réoriente si
- ΔR² (SSL vs scratch) between 0.05 and 0.15 OR shortcut flag (R²_probe ≥ 0.90 × R²_SSL) — pivot to richer pretext tasks (e.g., contrastive, diffusion-based) or multi-modal pretext (spectra + temperature/pH)
- Risques
- Synthetic generator does not capture real wastewater spectral complexity (e.g., non-linear scattering, matrix effects)Probabilité : highParade : Calibrate generator on multiple open datasets; validate against held-out real spectra; report domain gap; if gap too large, proceed to Phase 2 with real data
- Open UV-VIS time-series datasets are scarce or lack COD labelsProbabilité : mediumParade : Use synthetic data as primary; supplement with open datasets for qualitative validation; if no real data, Phase 1 remains purely synthetic and Phase 2 becomes critical
- Transformer architecture too small/large for synthetic data; overfitting or underfittingProbabilité : mediumParade : Hyperparameter search with Optuna; test d_model ∈ {32,64,128}, layers ∈ {2,4,6}; report sensitivity
- Temporal autocorrelation in synthetic data too strong/weak, biasing next-spectrum pretextProbabilité : mediumParade : Vary lag-1 ρ ∈ {0.5, 0.7, 0.85, 0.95}; report ΔR² as function of ρ
- Statistical power insufficient with 5 foldsProbabilité : lowParade : Increase folds to 10 or use repeated CV; run power analysis upfront