in silico
Phase 1 : In Silico Validation (Virtual Twin & Sensitivity Analysis)
- Objectif
- To falsify the core hypothesis using existing datasets and a virtual lake environment, before any new sediment cores are collected. This phase tests whether the proposed mechanism (Bayesian prior integration) can theoretically produce the predicted 20% CV reduction, and identifies the most sensitive parameters and data requirements.
- Coût estimé
- €1,500 - €3,000 (mostly computational time & personnel)
- Durée estimée
- 6-8 weeks
- Critères de réussite
- CV reduction from paleo-informed vs. contemporary-only calibration · ≥ 20% reduction (e.g., from 30% to ≤24%) in at least 4 out of 5 virtual lakes, even with proxy error inflation factor = 2.0 · (Paired t-test on log-transformed CV values from 1000 Monte Carlo simulations; report mean and 95% CI of the reduction.)
- Posterior narrowing of TP threshold parameter · ≥ 30% reduction in standard deviation of the posterior TP threshold distribution compared to the prior · (F-test for equality of variances (p < 0.05) comparing prior and posterior distributions.)
- Superiority over synthetic data augmentation · CV reduction from paleo-informed is ≥ 10 percentage points greater than from contemporary-augmented (e.g., 25% vs. 15%) · (Paired t-test comparing the two CV reduction distributions.)
- Power analysis · Power ≥ 0.80 to detect a 20% CV reduction with n ≤ 15 lakes · (Analytical power calculation using “pwr” package, based on observed effect size and inter-lake variance.)
- On continue si
- All four success criteria are met in the virtual environment. The mechanism is theoretically sound and the required sample size is feasible.
- On arrête si
- The CV reduction is < 10% in the majority of virtual lakes, or the required sample size exceeds 25 lakes. The hypothesis is likely false or the effect is too small to be practically significant.
- On réoriente si
- The CV reduction is 10-20%, or only met with proxy error inflation factor = 1.0. Pivot: Re-evaluate the proxy error assumptions, refine the prior construction method (e.g., use a different parameterization of the prior), or focus on a subset of lake types (e.g., only lowland lakes) where the effect is stronger.
- Risques
- PCLake model structure is too rigid to benefit from paleo priors (equifinality).Probabilité : mediumParade : Use a simpler, more flexible statistical model (e.g., a Bayesian hierarchical model) as a benchmark alongside PCLake. If PCLake fails, the statistical model may still validate the core mechanism.
- Available open-access datasets lack sufficient paired contemporary and paleo data.Probabilité : mediumParade : Expand the search to include regional datasets (e.g., from Europe via NERC, USGS). If data is still insufficient, simulate the paleo data from the model as planned, which is a valid approach for Phase 1.
- Computational cost of Bayesian calibration with PCLake is prohibitive.Probabilité : lowParade : Use emulators (e.g., Gaussian process surrogate models) to approximate PCLake outputs, or use a simplified version of PCLake (e.g., PCLake+ with reduced complexity).