Campaign-resolved evidence for adaptive critical-heat-flux surrogates under experimental shift

Technical Summary: Campaign-Resolved Evidence for Adaptive Critical-Heat-Flux Surrogates Under Experimental Shift

Problem Statement
Critical Heat Flux (CHF) is a limiting condition in water-cooled systems where heat removal deteriorates rapidly. While prediction methods have evolved from empirical correlations to physics-informed machine learning (ML), a fundamental gap remains: the reliability of models when applied to entirely new experimental campaigns. Existing validation often relies on record-level random splitting, which fails to account for the hierarchical structure of experimental data. Rows from the same campaign share specific test sections, instrumentation chains, surface histories, and operating trajectories. Consequently, a model trained on a subset of records from a campaign may not generalize to a completely unseen campaign, even if the thermohydraulic conditions appear similar. This study addresses the absence of a common evidentiary unit across the inference chain, treating the experimental campaign—not the individual database row—as the fundamental unit for generalization, uncertainty calibration, evidence valuation, and adaptation.

Methodology
The study utilizes a locked archive of 24,579 water-CHF measurements from 60 uniformly heated vertical round-tube campaigns. The methodology is built on a strict "campaign-held-out" protocol to prevent data leakage:

1. Campaign-Resolved Validation: The 60 campaigns are partitioned into 12 load-balanced outer blocks. In each block, one campaign is held out as the test set, while the remaining campaigns are split into training and validation sets. Preprocessing, model selection, ensemble weighting, and uncertainty calibration are determined exclusively from the training and validation campaigns.
2. Physics-Anchored Learning: Four candidate models are trained to predict the logarithmic residual of CHF relative to a reference. Three models learn residuals relative to the Bowring correlation (a standard engineering anchor), while one learns directly. The candidates include Extra Trees (record-weighted and campaign-balanced) and Histogram Gradient Boosting. A convex ensemble is formed using validation-only weights.
3. Uncertainty Quantification: Two conformal prediction rules are compared: a pooled-record rule and a "campaign-robust" rule that calculates quantiles across campaigns rather than records to ensure protection for typical held-out facilities.
4. Directed Transfer and Data Valuation: The study evaluates the value of transferring knowledge from source campaigns to target campaigns using three distinct "transport games" (Gaussian support, inverse distance, and ten-nearest campaigns). It calculates directed removal values and permutation Shapley values to determine if a source helps or harms a specific target, explicitly testing for negative transfer and directional asymmetry.
5. Sparse Adaptation: The framework tests whether revealing a small number of measurements (1, 3, 5, or 10) from a new target campaign allows for a regularized Bayesian update of the campaign-specific residual state. This is compared against various acquisition policies, including a complex "TWIN-VALUE" score, maximin coverage, and simple random sampling.
6. Independent Audit: An independent hierarchical analysis (TRACE) using radial-basis fields and random effects is performed to triangulate whether campaign-level structure persists across different estimators.
7. Thermal Verification: A one-dimensional energy balance is coupled to the surrogate to verify numerical conservation, explicitly distinguishing the work from a full multiphase digital twin.

Key Results

• Generalization Penalty: Whole-campaign prediction is significantly harder than record-level prediction. While the physics-anchored surrogate reduced record-weighted Mean Absolute Percentage Error (MAPE) from 9.67% to 4.52%, the campaign-macro MAPE (equal weight per campaign) was 7.47%. A weight-only audit revealed that historical sampling concentration explains only 39.2% of the gap between record and campaign metrics; the remaining 60.8% is attributed to residual campaign-difficulty heterogeneity (e.g., unrecorded surface states, metrology, or facility-specific biases).
• Directional Transfer: Cross-campaign evidence is not exchangeable. Of 3,540 directed source-to-target pairs, 50.9% were found to be harmful (negative transfer), while 43.4% were helpful. Crucially, transfer is directional: in 30% of reciprocal campaign pairs, a source that helps a target harms the reverse direction. Measured thermohydraulic distance and support explain only part of this structure.
• Uncertainty Calibration: Pooled-record calibration (91.85% coverage) failed to protect difficult campaigns, with lower-decile coverage dropping to 34.92%. The campaign-robust rule improved lower-decile coverage to 76.99% and campaign-macro coverage to 92.17%, but at the cost of wider prediction intervals (median width inflation of 1.47×).
• Sparse Adaptation: Revealing just five measurements from a new campaign reduced the macro MAPE from 7.47% to 5.65% (a 24.4% improvement). However, the complex TWIN-VALUE acquisition policy did not statistically outperform simple random sampling or maximin coverage. The primary benefit was the regularized state update itself, not the specific selection policy.
• Independent Triangulation: The TRACE audit confirmed that campaign effects are predominantly level-like (intercept shifts) rather than slope changes, and that a two-pattern response topology exists. This supports the interpretation that campaign identity acts as a reproducible, low-dimensional residual state.

Significance and Claims
The paper reframes CHF machine learning as an "experimental evidence transport problem" rather than a pure regression task. Its primary contribution is methodological and evidentiary:

1. Campaign Provenance is Critical: Experimental provenance retains a reproducible residual component that dictates how evidence transfers, how uncertainty is distributed, and how new experiments respond to sparse data.
2. Non-Exchangeability: Historical data value is directional and context-dependent. A broad, well-sampled campaign can support a narrow target, but the reverse may be harmful.
3. Adaptive Surrogates vs. Digital Twins: The study defines its output as an "offline, campaign-initialized adaptive surrogate." It explicitly rejects claims of operational digital twin status, mechanism identification, or multiphase simulation fidelity. The one-dimensional energy balance serves only as a conservation check.
4. Modest Acquisition Claims: While sparse updating is effective, the study fails to demonstrate that complex active learning policies outperform random sampling in this specific context. The "failed" acquisition-superiority gate is presented as a valid scientific result, emphasizing that the value lies in the state update, not the query strategy.
5. Falsifiability: The work concludes with a frozen, outcome-blind matrix of 30 conditions designed for future prospective falsification, acknowledging that current claims are retrospective and require independent experimental validation to establish external validity.

In summary, the study demonstrates that treating the experimental campaign as the unit of evidence reveals persistent heterogeneity that record-level analysis obscures. It provides a framework for deciding when historical evidence is transferable, when uncertainty must be widened, and when limited new measurements can support prediction updates, without claiming to solve the underlying physical mechanisms of the boiling crisis.