Source-linked AI summary
Amortized Localized Posterior Inference for Model Calibration with Application to an Ocean Turbulence Closure
Exaucé Luweh Adjim Ngarti, Élise Arnaud, Léo Nicoletti, Victor Trappler, Arthur Vidard
TL;DR
A single calibration parameter may not reproduce physical responses across contexts when model discrepancy varies, and uncertain contexts also affect calibration. The paper constructs localized posteriors and amortizes their approximation across contexts; its twin experiments report reduced predictive errors compared with global or deterministic calibration, within settings designed to admit local compensation.
Problem
A single calibration parameter may not reproduce responses across contexts when model discrepancy varies, while uncertainty in observations and contexts can affect inferred parameters and predictions.
Source: §1 · p. 1, §2.1 · p. 3Method
The method weights observation likelihoods by proximity to a target context, adjusts localization for context uncertainty, and uses a conditional flow to approximate the posterior family.
Source: §1 · pp. 2–3, §8 · p. 22Results
Localized calibration reduces temperature and zonal-velocity errors relative to deterministic and global-posterior calibration in the ocean-model twin experiment.
Source: §8 · p. 22Takeaways & Limitations
In the evaluated ocean experiment, localized calibration adapts across imposed forcing regimes, but parameter identifiability is reduced under weak-wind heating conditions.
Source: §8 · p. 22, §7 · p. 22Applicability requires local compensability and sufficient information from nearby contexts; sparse sampling, large context uncertainty, or rapid parameter changes restrict useful localization strength.
Source: §7 · p. 22
Abstract
from arXiv · showhide
Model discrepancy may prevent a single calibration parameter from reproducing a physical response across the context domain. In addition, uncertainties in observations and contexts affect calibration and must be accounted for and quantified. We construct a family of localized posteriors by weighting observation likelihoods according to their proximity to a target context. Gaussian localization weights incorporate context uncertainty through covariance inflation, with reflection at domain boundaries and an optional correction for uneven context sampling. A conditional normalizing flow approximates the family of posteriors through amortized variational inference. Integrating the localized posteriors under a prescribed context measure gives the marginalized parameter distribution. In a twin experiment with the Branin-Hoo model, we compare global and localized inference under uniform and skewed context distributions. Localization reduces the reported aggregate predictive errors relative to the global posterior. An application to the k-$\epsilon$ turbulence closure in a single-column ocean model, with atmospheric surface forcing as context, shows that localized posterior calibration reduces temperature and zonal-velocity errors relative to deterministic and global-posterior calibration. Both twin experiments are constructed to admit local compensation.
1. Introduction
Calibration can fail to represent physical responses across contexts when model discrepancy varies, and uncertain or unevenly sampled contexts can affect inference. The paper addresses this with localized, amortized posteriors and reports lower predictive errors than global or deterministic calibration in its experiments.
1. Introduction
- A single calibration parameter may not explain observations across contexts when simplified or missing mechanisms cause context-dependent model discrepancy, while context sampling and recorded contexts may also be uncertain.These factors can affect inferred parameters and predictive performance across the context domain.Source: §1 · pp. 1–2
- The method constructs a family of localized posteriors by weighting observation likelihoods according to proximity to a target context, under a local compensability assumption.Observations near the target context receive greater weight, allowing the calibration distribution to vary across contexts.Source: §1 · p. 2
- Gaussian localization weights incorporate context uncertainty through covariance inflation, with reflection at domain boundaries and an optional correction for uneven context sampling.Integrating localized posteriors under a prescribed context measure yields a marginalized parameter distribution.Source: §1 · pp. 2–3
- A conditional normalizing flow trained with amortized variational inference approximates localized posteriors across contexts without repeating optimization for each target.The flow can be evaluated at a new target context.Source: §1 · pp. 2–3
- In the Branin–Hoo twin experiment, localization reduces aggregate predictive errors relative to global posterior calibration under uniform and skewed context distributions.The experiment is constructed so local compensation is possible.Source: §1 · p. 2
- Localized posterior calibration reduces temperature and zonal-velocity errors relative to deterministic and global-posterior calibration in the ocean-model experiment.The experiment uses atmospheric surface forcing as context and also finds reduced parameter identifiability under weak-wind and heating conditions.Source: §1 · p. 2
2. Problem setting
The problem setting treats physical responses as numerical-model outputs plus context-dependent discrepancy and frames calibration around uncertain observations and recorded contexts. Its objective is local calibration across contexts and parameter uncertainty relevant to a prescribed context measure.
2.1. Model discrepancy and context
- Because discrepancy may vary with context, a single parameter need not fit every context; local compensability assumes a suitable parameter can approximately compensate nearby.The approach also assumes adequate parameters at nearby contexts are sufficiently similar for nearby observations to be informative.Source: §2.1 · p. 3
2.2. Observation and context uncertainty
- The working likelihood assumes conditionally independent observations with additive Gaussian errors, while recorded contexts have independent additive Gaussian errors with known covariance.Context uncertainty enters localization weights rather than an explicit integration of the model over unknown true contexts.Source: §2.2 · pp. 3–4
2. Problem setting
- Calibration uses observations collected under different physical contexts, with the goal of adequate local calibration across the context domain.The objective accounts for observation and context uncertainty and uneven context sampling.Source: §2 · p. 3, §2.3 · p. 4
2.3. Calibration objectives
- The calibration target is a parameter distribution relevant to a prescribed context probability measure, which need not match the dataset’s empirical context distribution.The resulting context-dependent posterior family is intended for approximation and aggregation.Source: §2.3 · p. 4
3. Amortized localized posterior inference
The method builds context-specific posteriors by weighting observations near each target context, then amortizes their approximation and aggregates them under a prescribed context measure. Localization accounts for context uncertainty, domain boundaries, and optionally uneven context sampling.
3.1. Localized posterior
- Localized posteriors give greater weight to observations near the target context, with a scaling factor based on the weights’ effective sample size.The scaling factor adjusts the weighted likelihood’s contribution relative to the prior.Source: §3.1 · p. 4, §3.3 · p. 6
3.2. Localization weights
- Gaussian localization weights incorporate context uncertainty through covariance inflation, reflect at domain boundaries, and can optionally correct for uneven context sampling.The optional sampling correction uses an estimated or known context sampling density.Source: §3.2 · p. 5
3.3. Posterior scaling
- The localized-posterior scaling factor equals the Kish effective sample size, ranging from n for uniform weights to one when a single observation receives all weight.These cases recover the full-data log-likelihood and a single observation’s likelihood contribution, respectively.Source: §3.3 · p. 6
3.4. Bandwidth selection principle
- A localization rate expresses weight concentration as a percentage, and a bandwidth is selected separately for each target context to meet the prescribed rate.The prescribed rate is selected using predictive performance metrics, and bandwidth computation uses Brent’s method when a root is bracketed.Source: §3.4 · p. 6
3.5. Conditional flow and variational objective
- A conditional normalizing flow is trained by minimizing the average negative ELBO across target contexts to approximate the localized-posterior family without repeating inference at each context.Observation indices sampled according to localization weights provide an unbiased Monte Carlo estimate of the weighted log-likelihood sum.Source: §3.5 · pp. 6–7
3.6. Aggregation under a prescribed context measure
- Integrating the conditional-flow posteriors under the prescribed context measure produces a mixture weighted by that measure, approximated by sampling contexts and corresponding parameter draws.The distribution summarizes calibration uncertainty for the contexts specified by Pctx.Source: §3.6 · pp. 7–8
4. Reference methods and predictive evaluation
The evaluation compares localized calibration with global-posterior references using point and ensemble predictions. Twin experiments provide noise-free physical responses for evaluating predictions made at recorded noisy contexts.
4.1. Reference methods
- The global-posterior references are MAP-based deterministic calibration and a normalizing-flow variational approximation using all observations without localization.Point predictions use either the MAP estimate or componentwise posterior median; ensemble predictions propagate posterior parameter samples to quantify predictive uncertainty.Source: §4.1 · p. 8
4.2. Point and ensemble errors
- Predictions at recorded noisy contexts are compared with noise-free twin-experiment responses using absolute and relative discrepancies for point and ensemble predictions.For absolute discrepancy, the ensemble criterion is the empirical energy score scaled by 1/d_y, reducing to CRPS in one dimension.Source: §4.2 · p. 8
4.3. Aggregate metrics
- Across test contexts, aggregate evaluation reports mean error, empirical standard deviation, and upper-tail mean error, with the upper tail defined above the empirical 0.9 quantile.The upper-tail mean measures the average of errors above that quantile.Source: §4.3 · pp. 8–9
5. Branin–Hoo twin experiment
The Branin–Hoo twin experiment tests localized calibration where the synthetic physical response changes between parameter regimes. Localization lowers aggregate predictive errors in reported comparisons, while results vary with localization rate and context sampling.
5.1. Model, reference response, and data
- The Branin–Hoo twin experiment uses a synthetic response interpolating between two parameter regimes and uniform or Beta (2, 3) calibration contexts.The transition prevents a single calibration parameter from representing both reference regimes; observation-noise variance is 52 and context-noise variance is 0.052.Source: §5 · p. 9, §5.1 · pp. 9–10
5.2. Inference settings
- All methods used Adam for 1,500 iterations; global and conditional flows used three spline coupling layers, and the conditional network had two hidden layers of width 32.Variational training used 16 particles and a target-context batch size of 64.Source: §5.2 · p. 10
5.3. Sensitivity to localization rate
- With skewed calibration, stronger localization decreases errors on uniform test contexts but can increase them on skewed tests; with uniform calibration, metrics decrease for both test distributions.The skewed test distribution emphasizes the lower-context regime, whereas uniform evaluation covers the transition and second regime.Source: §5.3 · p. 10
5.4. Aggregate ensemble performance
- Density correction's effect depends on localization rate and test distribution, improving some reported means while worsening others.For skewed calibration and uniform testing, the mean changes from 15.382 to 8.525 at 30%, but from 7.169 to 7.334 at 70%.Source: §5.4 · p. 10
5.5. Predictions and marginalized parameter distributions
- The posterior predictive medians and 90% credible intervals show localized inference following the changing physical response more closely than a single global calibration.Figure 3 compares localization rates of 30% and 70% under skewed and uniform calibration.Source: §5.5 · p. 10, Figure 3 · p. 13
- Marginalized parameter distributions change with the prescribed context measure, while skewed calibration sampling continues to influence their structure.Uniform calibration gives clearer separation between contributions from the two parameter regimes.Source: §5.5 · p. 10, Figure 4 · p. 14
5.4. Aggregate ensemble performance
- Mean absolute ensemble energy score fell from 43.537 ± 19.324 for the global posterior to 6.679 ± 0.265 at 30% and 6.699±0.222 at 70% localization under uniform calibration and testing.Localized methods also reduce reported upper-tail errors relative to the global posterior.Source: §5.4 · p. 10, Table 1 · p. 12
6. Ocean turbulence-closure twin experiment
The ocean twin experiment calibrates two k–ε closure parameters across uncertain atmospheric forcing contexts, with reference parameters varying by forcing regime. Localized calibration lowers aggregate predictive errors, while posterior summaries and profiles show regime-dependent behavior.
6.1. Single-column model, transport, and surface forcing
- The single-column model evolves zonal velocity, temperature, and salinity, with wind stress and surface heat flux defining the forcing context.Temperature and zonal velocity are the observed variables; vertical turbulent diffusivity and viscosity govern transport.Source: §6.1 · p. 15
6.2. Turbulence closure and calibrated parameters
- Calibration targets cµ0 and cε1: cε1 weights shear production, while cµ0 relates the turbulent length scale to k and ε.The calibrated parameter bounds are cµ0 ∈ [0.3, 0.8] and cε1 ∈ [0.1, 1.9].Source: §6.2 · p. 15
6.3. Forcing-dependent reference parameters
- The twin reference uses context-dependent parameter distortion: wind forcing changes cε1, and heat forcing changes cµ0, producing four forcing regimes.The imposed distortion is illustrated in Fig. 5.Source: §6.3 · p. 16, Figure 5 · p. 16
6.4. Data, uncertainty, and numerical configuration
- The experiment uses 1,024 calibration and 512 test contexts, with temperature and zonal velocity recorded over 100 depth cells and 73 times.Observation noise is set to 5% of each reference-signal standard deviation, and recorded contexts have independent Gaussian errors.Source: §6.4 · p. 16
6.5. Inference and predictive evaluation
- Aggregate predictive metrics are evaluated separately for temperature and zonal velocity over 512 test contexts, using predictions at recorded noisy contexts.Profile comparisons include posterior predictive medians, 90% intervals, the twin reference, and deterministic MAP predictions.Source: §6.5 · p. 17
6.6. Aggregate errors and parameter distributions
- The marginalized localized posterior has three modes for weak-wind cooling, strong-wind cooling, and strong-wind heating; weak-wind heating is less identifiable.That regime has limited mixing and lower response sensitivity to closure parameters; Fig. 6 uses 32 samples at each of 1,000 contexts and 98% localization.Source: §6.6 · p. 17, Figure 6 · p. 19
- At 98% localization, localized-posterior median predictions yield mean RMSEs of 0.010 97 K for temperature and 0.025 30 m/s for zonal velocity, while ensemble scores are 0.010 44 K and 0.023 05 m/s.Source: §6.6 · p. 17
6.7. Time-averaged profiles across forcing regimes
- Figures 7 and 8 compare reference, localized and global posterior-median, and MAP time-averaged temperature and zonal-velocity profiles across representative forcing regimes.The figures show the upper mixed-layer depth range and use 98% localization; shaded regions indicate 90% credible intervals.Source: Figure 7 · p. 20, Figure 8 · p. 21
7. Discussion
The discussion finds that localization reduces predictive errors in both twin experiments, while the ocean experiment tests imposed parameter dependence rather than discrepancy from a missing mechanism. Its usefulness is bounded by local compensability, nearby-context information, and approximation error.
7. Discussion
- In the Branin–Hoo experiment, localization reduces predictive errors, but higher localization can improve full-domain performance while worsening performance under a test distribution concentrated in one regime.Density correction shifts this balance but does not remove dependence on the calibration context distribution.Source: §7 · p. 22
- In the ocean experiment, localized inference reduces aggregate errors and yields context-varying intervals, but the reference is generated with context-dependent parameters by construction.The experiment therefore tests imposed parameter dependence, not discrepancy caused by a missing physical mechanism; weak-wind heating also remains less identifiable when mixing sensitivity is low.Source: §7 · p. 22
- Useful localization requires local compensability and sufficient information from nearby contexts; sparse sampling, large context uncertainty, or rapid parameter changes restrict its strength.The approach preserves the original numerical model and parameter interpretation.Source: §7 · p. 22
- Conditional flows amortize inference across contexts, but flow approximation and optimization remain additional sources of error.Source: §7 · p. 22
8. Conclusion
Localized calibration reduces aggregate predictive errors in the Branin–Hoo experiment and improves temperature and zonal-velocity predictions in the ocean-model twin experiment. The method approximates context-specific posteriors and aggregates them under a prescribed context measure.
8. Conclusion
- The method uses weighted likelihoods, effective-sample-size scaling, and context-noise adjustment; a conditional normalizing flow approximates the posterior family for aggregation under a prescribed context measure.The localized calibration relies on approximate local compensation of model discrepancy.Source: §8 · p. 22
- Localization reduces aggregate predictive errors in Branin–Hoo and improves temperature and zonal-velocity predictions in the ocean-model twin experiment.Both evaluated localization rates reduce point and ensemble error aggregates relative to deterministic and global-posterior calibration.Source: §8 · p. 22
A.1. Localized ELBO identity
- The localized ELBO is a lower bound on the log normalization constant, so maximizing it is equivalent to minimizing the negative ELBO.Its identity expresses the localized ELBO through the normalization constant and a KL divergence.Source: §A.1 · p. 23
A.2. Localization precomputation
- Bandwidth precomputation expands an upper bracket and applies Brent’s method to match a prescribed localization rate; it cannot determine a bandwidth if the rate cannot be bracketed.The rate is computed from final normalized weights, including selected corrections.Source: §A.2 · p. 23
A.3. Conditional-flow training
The conditional flow is trained with a stochastic negative ELBO using localized observation sampling. The appendices describe kernel-weight behavior across context sampling schemes and examine point-prediction error across localization rates and aggregate metrics.
A.3. Conditional-flow training
- For each target mini-batch, parameter particles come from the conditional flow and observation indices are sampled using precomputed localization weights to train with the stochastic negative ELBO.The procedure differentiates through reparameterized particles and updates the flow parameters.Source: §A.3 · p. 24
B.1. Uniform context sampling
- Figure 9 compares average localization weights under uniform context sampling for Gaussian and reflected Gaussian kernels against the nominal uniform weight.Observed and target contexts are sampled uniformly on the context domain.Source: Figure 9 · p. 25, §B.1 · p. 25
B.2. Skewed context sampling
- Under skewed context sampling, Figure 10 compares reflected Gaussian weights with and without the density correction.The plotted average is taken across observed contexts when target contexts are sampled from the same domain.Source: Figure 10 · p. 25, §B.2 · p. 25
C.1. Point-prediction sensitivity to localization rate
- Figures 11 and 12 show MAP-based and posterior-median point-prediction error variation across localization rates for skewed and uniform calibration contexts.The first row corresponds to skewed data and the second to uniform data.Source: Figure 11 · p. 26, §C.1 · pp. 26–27, Figure 12 · p. 27
C.2. Aggregate point-prediction errors
- Table 3 reports posterior-median and mode point-prediction RMSE aggregates as mean error, test-context standard deviation, and 0.9 upper-tail mean.Each cell is reported as mean ± one standard deviation over the reported seed runs.Source: Table 3 · p. 28
C.3. Context-wise point errors and relative ensemble diagnostic
The Branin–Hoo diagnostics compare context-wise point errors and ensemble error across calibration distributions and selected localization rates, while marginalized posteriors are shown under alternative context measures. The ocean-model setup couples state, forcing, and turbulence closure, with smooth transitions used to define forcing regimes.
C.3. Context-wise point errors and relative ensemble diagnostic
- Across contexts, relative ensemble errors concentrate differently around the designed transition and the less represented regime.The diagnostic compares skewed and uniform calibration and selected localization rates.Source: §C.3 · p. 29, Figure 15 · p. 31
- Figures 13 and 14 compare context-wise MAP and posterior-median point-prediction errors across skewed and uniform calibration and selected localization rates.Rows represent calibration context distributions and columns represent localization rates.Source: Figure 13 · p. 29, Figure 14 · p. 30
C.4. Marginalized posteriors at 30% localization
- At 30% localization, marginalized posteriors are shown for three context measures, with reference parameter values marking the two regimes.The measures concentrate toward smaller contexts, are uniform, or concentrate toward larger contexts.Source: Figure 16 · p. 32, §C.4 · p. 32
D.1. Single-column configuration
- Table 4 presents single-column model configuration parameters.Source: Table 4 · p. 33
D.2. Turbulence-closure parameter settings
- Table 5 presents the default parameters of the k–ε turbulence closure.Source: Table 5 · p. 34
D.3. Smooth distortion and forcing regimes
- The forcing-regime setup uses a quintic smooth-transition function, illustrated for three values of δ.Table 6 summarizes four main parameter-distortion regimes.Source: Table 6 · p. 34, Figure 17 · p. 35
D.4. State–closure coupling and mixed-layer diagnostic
- The single-column state determines local shear and stratification for the k–ε closure, which returns turbulent diffusivity KT and viscosity KM to close state evolution.Temperature and zonal velocity are synthetic calibration observations; profile diagnostics use the mixed-layer depth method of.Source: §D.4 · p. 35, Figure 18 · p. 36
E.1. Data and prior
The calibration prior is uniform over specified ranges for the two closure coefficients, and the appendix reports method settings and final-time ocean profiles. The profiles compare localized, global, MAP, and reference predictions at 72 hours.
E.1. Data and prior
- The prior is uniform on Θ = [0.3, 0.8] × [0.1, 1.9] for (cµ0, cε1).Table 7 summarizes the calibration configuration.Source: §E.1 · p. 37, Table 7 · p. 37
E.2. Optimization, flow architecture, and bandwidth settings
- Bandwidth precomputation uses a reflected Gaussian kernel, minimum bandwidth 10−6, initial bracket distance multiplier 4, expansion factor 2, and at most 100 expansions.The target rate must be attainable and bracketed for root finding; Table 8 describes the MAP, global-flow, and localized-flow configurations.Source: §E.2 · p. 37, Table 8 · p. 38
F.1. Temperature
- At 72 hours, Figure 19 compares final temperature profiles against reference twin-simulation profiles using localized and global posterior medians and MAP predictions.Localized curves use 98% localization; shaded regions show 90% credible intervals.Source: Figure 19 · p. 39, Appendix F · p. 39
F.2. Zonal velocity
- At 72 hours, Figure 20 compares final zonal-velocity profiles against reference twin-simulation profiles using localized and global posterior medians and MAP predictions.Localized curves use 98% localization; shaded regions show 90% credible intervals.Source: Appendix F · p. 39, Figure 20 · p. 40