PaperlayerView exact revision on arXiv

Source-linked AI summary

A Data-Efficient Analytical Prior Machine Learning Framework for Sound Reduction Frequency Prediction in Helmholtz Resonators

Jiaming Li

arXiv:2608.16873v2cs.LG

TL;DR

Scarce high-fidelity simulation data limit purely data-driven prediction for resonators, even when a validated analytical model captures much of the response. This framework reuses that model through explicit discrepancy correction or distilled-prior calibration, improving data efficiency across limited simulation budgets.

  • Problem

    Expensive finite-element labels limit data-driven surrogate accuracy, while validated analytical models already capture part of the dominant resonator response.

  • Method

    The framework learns analytical-to-simulation discrepancies when the analytical model remains available, or distills and calibrates an analytical prior for self-contained prediction.

  • Results

    Across 20–70 simulation-labelled cases, both routes improved data efficiency; residual SVR reduced MAE to 0.426 Hz, while fine-tuned prior MLP reached 0.371 Hz.

  • Takeaways & Limitations

    Analytical or other low-cost engineering models can provide reusable prior information for data-efficient high-fidelity surrogate prediction under scarce-data conditions.

  • Takeaways & Limitations

    Accuracy was assessed against COMSOL simulations rather than laboratory measurements, so experimental validation is needed under more realistic conditions.

Abstract

from arXiv · show

High-fidelity finite-element simulations can provide accurate numerical predictions for side-branch resonators, but large simulation datasets are expensive to generate and purely data-driven surrogates may become unreliable when simulation-labelled data are scarce. This study develops an analytical-prior learning framework that reuses a low-cost analytical model to improve data efficiency under limited high-fidelity simulation budgets. Two complementary routes are considered. When the analytical model remains available at inference, it is retained as an explicit baseline and the simulation data are used to learn only the analytical-to-simulation discrepancy. When a self-contained predictor is required, the analytical mapping is first distilled from abundant low-cost evaluations into a learned prior and then calibrated with the limited simulation data. The framework is evaluated on rectangular side-branch Helmholtz resonators using 86 simulation-labelled geometries and 8,998 non-overlapping analytical-only geometries. The analytical model achieved a mean absolute error (MAE) of 1.333 Hz. Direct support vector regression (SVR) achieved 3.375 Hz, while residual SVR reduced the MAE to 0.426 Hz. A direct multilayer perceptron (MLP) achieved 1.109 Hz, whereas analytical-prior pretraining reduced the error to 0.556 Hz with frozen-prior residual adaptation and 0.371 Hz with full-model fine-tuning. Across training budgets of 20 to 70 simulation-labelled cases, both analytical correction and analytical-prior pretraining consistently improved data efficiency relative to direct learning. These results show that analytical prior information can substantially improve high-fidelity prediction when simulation data are scarce, with explicit correction and prior distillation serving complementary deployment needs.

1. Introduction

The study proposes an analytical-prior learning framework to improve data efficiency when high-fidelity simulation data are scarce. It evaluates explicit analytical correction and analytical-prior distillation as complementary routes for rectangular side-branch Helmholtz resonators.

  • Motivation: Analytical models provide rapid, physically interpretable frequency predictions, whereas finite-element simulations capture geometric and wave effects with greater fidelity but require substantially more computation.This fidelity–cost trade-off motivates combining analytical and simulation-based information.
  • Motivation: Conventional data-driven surrogates learn the complete geometry-to-response mapping, making accuracy dependent on expensive simulation-label quantity and coverage.With limited simulation budgets, models must learn both dominant physical trends and high-fidelity corrections.
  • Framework: The framework uses two routes: explicit analytical correction when the analytical model remains available, and analytical-prior distillation when a self-contained predictor is required.The explicit route learns only the analytical-to-simulation discrepancy, while the distilled route calibrates a learned prior with limited high-fidelity data.
  • Evaluation: The evaluation combines 8,998 analytical-only geometries with 86 simulation-labelled rectangular side-branch Helmholtz resonators.The simulation-labelled cases contain paired analytical and COMSOL frequencies.
  • Evaluation: Residual support vector regression evaluates the explicit-prior route, while compact multilayer perceptrons evaluate analytical-prior distillation and subsequent high-fidelity calibration.The study is framed within multi-fidelity learning, information fusion, transfer learning, residual correction, and physics-guided machine learning.

2. Methodology

The framework uses scarce high-fidelity simulation data to learn information not represented by an informative but imperfect analytical model. It provides explicit analytical correction when the analytical model remains available and analytical-prior distillation when a self-contained predictor is required.

  • Data construction: Each simulation-labelled case pairs design variables with analytical and high-fidelity responses, whereas the analytical-only dataset contains analytical evaluations.These paired and analytical-only datasets support distinct uses of the available information.
  • Framework assumptions: The framework assumes shared inputs and targets across fidelities, with many more low-cost analytical evaluations than high-fidelity simulations.The intended regime is N_A≫N_S, and the analytical model is informative but imperfect.
  • Direct learner: Direct learning uses only scarce simulation-labelled pairs and ignores the analytical prediction, providing the high-fidelity-only reference.This reference places the complete geometry-to-high-fidelity mapping on the limited simulation data.
  • Explicit analytical correction: Explicit analytical correction models the analytical-to-simulation discrepancy while retaining the analytical response as the prediction baseline.This route focuses simulation-labelled data on the remaining correction but requires analytical evaluation for every new input at inference.
  • Analytical-prior distillation: Analytical-prior distillation first learns the analytical mapping from analytical-only data, then calibrates it with limited simulations by fine-tuning all parameters or adapting a frozen prior residually.Both calibrated predictors are self-contained; full-model fine-tuning permits global adaptation, whereas frozen-prior residual adaptation preserves the distilled mapping.

3. Case-study validation

The case study validates analytical-prior learning for sound-reduction frequency prediction in rectangular side-branch Helmholtz resonators using paired analytical and COMSOL data. It compares direct learning, explicit analytical correction, and distilled-prior calibration under matched high-fidelity data budgets.

  • Case-study design: The validation family varies cavity dimensions, neck geometry, and cavity asymmetry across rectangular side-branch Helmholtz resonators.The prediction target is each resonator geometry’s sound-reduction frequency.
  • Explicit-prior route: The explicit-prior route retains A(x) during inference and uses residual SVR to predict the discrepancy S(x) − A(x), whereas direct SVR predicts S(x) from geometric variables.Both SVR formulations use identical outer folds, radial-basis-function kernels, and standardized input features.
  • Datasets: 86 simulation-labelled geometries are paired with 8,998 non-overlapping analytical-only cases after removing two exact geometry matches.The simulation-labelled set contains 21 symmetric single-neck, 24 asymmetric single-neck, and 41 multiple-opening cases.
  • Distilled-prior route: The distilled-prior route learns P(x) ≈ A(x) from 8,998 analytical-only cases, then calibrates it with 86 simulation-labelled cases using full-model fine-tuning or frozen-prior residual adaptation.The same repeated stratified fivefold outer partitions with five repeats are used across simulation-based comparisons.

4. Results

Analytical-prior methods improved prediction accuracy over matched direct-learning references in both explicit-correction and distilled-prior routes. These gains persisted across reduced simulation-labelled training budgets, although the cross-model advantage between the two best methods was not decisive.

  • Explicit-prior route: 68.0% lower MAE than the analytical model and 87.4% lower than direct SVR were achieved by residual SVR.Residual SVR was compared with the analytical model and direct SVR as matched references.
  • Distilled-prior route: 49.8% and 66.6% MAE reductions were achieved by frozen-prior residual MLP and full-model calibrated MLP, respectively, versus direct MLP; the latter reached 0.371 Hz and R² = 0.998.Full-model calibrated MLP gave the lowest numerical MAE in the distilled-prior comparison.
  • Statistical comparison: 0.055 Hz separated full-model calibrated MLP and residual SVR, but the 95% bootstrap interval crossed zero and the comparison was not decisive.Full-model fine-tuning outperformed frozen-prior residual adaptation and had lower MAE in 87.5% of bootstrap resamples.
  • Prior distillation: 0.499 Hz MAE and R² = 0.997 measured the distilled prior against the analytical prediction, while its COMSOL-target MAE was 1.342 Hz versus 1.333 Hz for the exact analytical model.The prior was evaluated on the 86 simulation geometries excluded from analytical pretraining.
  • Small-data robustness: 0.786 versus 4.448 Hz at 20 cases and 0.445 versus 3.427 Hz at 70 cases were the residual-SVR and direct-SVR mean test MAEs.The distilled-prior comparison showed the same pattern: full-model calibrated MLP achieved 0.846 versus 3.938 Hz at 20 cases and 0.339 versus 1.105 Hz at 70 cases; comparisons were matched within budget.

5. Discussion

The discussion finds that incorporating low-cost analytical information improves learning from scarce high-fidelity simulations through two complementary deployment routes. It also identifies route-specific tradeoffs and limits the conclusions to the evaluated simulation setting, geometry family, target, and architectures.

  • Analytical-prior benefit: Analytical information improves data efficiency because it captures the dominant geometry-frequency relationship, leaving learners to model mainly the analytical-to-simulation discrepancy.Direct learners must reconstruct both the broad trend and the remaining discrepancy from limited simulation-labelled data.
  • Analytical-prior benefit: Residual SVR outperformed direct SVR, while both distilled-prior MLP calibrations improved on direct MLP across budgets from 70 to 20 simulation cases.The controlled comparisons attribute the benefit to how analytical information is used rather than to additional high-fidelity data.
  • Deployment routes: Explicit analytical correction suits inference workflows retaining the analytical model, whereas analytical-prior distillation suits self-contained predictors learned from inexpensive analytical evaluations.Residual SVR retains the exact analytical prediction as its baseline; distillation first learns the analytical mapping before simulation calibration.
  • Deployment routes: Full-model fine-tuning achieved lower error than frozen-prior residual adaptation because it can adjust the complete pretrained representation toward the simulation response.The frozen route preserves the distilled prior and restricts high-fidelity learning to an additive correction that also absorbs distillation error.
  • Deployment routes: Residual SVR and full-model calibrated MLP had numerically close errors, so route selection should reflect inference workflow as well as predictive accuracy.Residual SVR preserves an explicit analytical component, whereas the calibrated MLP produces a single geometry-to-high-fidelity predictor.
  • Limitations: The findings are limited by simulation-only validation, 86 geometries from one rectangular resonator family, a single predicted frequency, and a limited architecture set.Future work should add experimental validation, new geometries, applicability-domain assessment, spectra and inverse-design tasks, uncertainty-aware prediction, and external testing.

6. Conclusion

The study develops an analytical-prior framework with explicit discrepancy correction and distilled-prior calibration routes for prediction when high-fidelity data are scarce. Validation on Helmholtz resonators shows improved accuracy and data efficiency relative to direct learning.

  • The framework reuses a low-cost analytical model either as an explicit inference-time baseline or as a distilled prior for a self-contained predictor.In the explicit route, scarce high-fidelity data learn only the analytical-to-simulation discrepancy; in the distilled route, the analytical mapping is learned first and then calibrated.
  • Residual SVR achieved 0.426 Hz MAE, compared with 1.333 Hz for the analytical model and 3.375 Hz for direct SVR.The validation used 8,998 analytical-only geometries and 86 simulation-labelled geometries.
  • Full-model fine-tuning achieved 0.371 Hz MAE, while frozen-prior residual adaptation achieved 0.556 Hz and direct MLP achieved 1.109 Hz.
  • The results show that analytical-prior routes use scarce high-fidelity data more efficiently for refinement and calibration than for reconstructing the complete mapping from scratch.
  • The approach extends to analytical, reduced-order, or empirical engineering models as reusable prior information for data-efficient high-fidelity prediction.Explicit correction and prior distillation serve complementary deployment needs.
Loading 2608.16873v2…