acceptodds
Under review as a conference paper at ICLR 2027

AutoMF: Automated Ensembles for Multifidelity Field Prediction

Abstract

Learning to predict physical fields requires fine simulation data that can be expensive to obtain. Multifidelity learning combines coarse and fine solutions, but choosing a surrogate from limited fine data can discard useful predictions. We introduce AutoMF, an automated ensemble workflow built around nine surrogates adapted from neural operators, convolutional networks, and reduced basis models. The library includes transfer learning, pooled fidelity training, and coarse field correction, together with a fine only reference. We assemble 22 datasets spanning seven classes of partial differential equations (PDEs) and climate emulation to evaluate whether combining predictions reduces errors in model choice. Five additional input and fine solution pairs, reserved from training, determine one weight per model using established ensemble rules. Weights stay fixed across the field and new inputs, which require no additional coarse simulation. The fitted mixture achieves lower error on 17 of 22 datasets than selecting one model from the same fitting examples. It reduces geometric mean relative error by 11.6% under equal weighting of the seven PDE classes and by 7.5% under equal weighting of all 22 datasets. Against the best implemented baseline chosen separately for each PDE dataset after evaluation, its reduction is 29.4% under equal weighting of the seven PDE classes. Code and reproducibility materials are available in an [anonymous repository](https://anonymous.4open.science/r/mf_field-35BF/README.md).

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
multifidelity learning, surrogate modeling, scientific machine learning, ensemble learning, neural operators, field prediction

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_SC62OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces AutoMF, an ensemble workflow for multifidelity field prediction built around nine adapted surrogate models, including neural operators, convolutional networks, and reduced-basis models. For each dataset, five reserved high-fidelity examples determine one scalar weight per model, shared across inputs and spatial locations. The authors compare model selection, inverse-error weighting, and jointly optimized ensemble weights. The fitted ensemble achieves lower evaluation error than selecting a single model using the same fitting examples on 17 of 22 datasets. The paper provides detailed comparisons of individual surrogates and ensemble strategies across the benchmark.

Critical strengths

The evaluation includes a diverse collection of datasets and model types, providing evidence for the workflow’s usefulness across a broad range of tested multifidelity prediction settings. The variation in individual model performance across tasks also motivates considering multiple surrogate approaches rather than relying on one architecture throughout the benchmark.

The detailed performance comparisons and field visualizations help readers understand the benefits and limitations of combining predictions. In particular, the visualization of predicted fields and spatial errors illustrates how an ensemble can improve on individual models in a successful example. Comparisons with simpler weighting rules and smaller model subsets further help assess which aspects of the workflow contribute to its gains.

Critical weaknesses

The paper does not fully quantify the additional training time, memory requirements, and inference costs associated with maintaining and evaluating multiple surrogate models. Although the accuracy gains are demonstrated, a more complete computational comparison would help establish whether these gains justify the additional resources. This is particularly relevant because selecting a single model and constructing an ensemble have different deployment costs, even when both begin with the same trained library.

The paper includes useful comparisons with inverse-error weighting, uniform averaging, and smaller fitted ensembles. However, these controls largely retain the same adapted surrogate library. Comparisons with ensembles constructed from the implemented baseline models, or repeated training runs of a single architecture, would help isolate the benefit of this particular library. Evaluating such alternatives under comparable computational budgets would strengthen the practical conclusions.

Questions for authors

1.Beyond the reported weighting and subset controls, have you evaluated ensembles constructed from the implemented baseline models or repeated training runs of a single architecture? If such results are available, how do they compare with AutoMF?

2.Can you provide computational measurements, such as total training time, inference latency, and memory requirements, to contextualize the accuracy improvements over selecting a single model?

3.For cases such as Poisson II, where the fitted mixture performs worse than the selected model, what weights are assigned to the individual models? Does this deterioration appear only on evaluation examples, or also on the fitting examples when measured using the reported mean relative L2L_2 error? This would help distinguish generalization effects from the difference between the fitting and reporting objectives.

Reviewer_JNn2OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes AutoMF, an automated ensemble for multifidelity field prediction. It trains a heterogeneous library of nine surrogate models, including FNO-, ConvNeXt-, DeepONet-, wavelet-, correction-, and POD-GP-based models, and uses a small number of additional high-fidelity examples to either select one model or learn ensemble weights. Across 22 PDE/climate datasets, the fitted ensemble beats model selection on 17/22 tasks and reduces aggregate error by 11.6% under PDE-class-balanced evaluation

Critical strengths
  1. The paper evaluates nine surrogate models over 22 datasets spanning seven PDE classes plus ERA5 climate emulation, which is substantially broader than a typical single-PDE evaluation.
  2. Only a small number of reserved fine examples are needed to estimate the ensemble weights, and the weights remain fixed for new inputs; importantly, deployment does not require a new coarse simulation
  3. The authors compare model selection, inverse-error weighting, fitted weighting, uniform averaging, smaller subsets, and multiple fitting budgets. They also provide detailed reproducibility material, saved predictions, fixed splits, and replay scripts.
Critical weaknesses
  1. The ensemble rules themselves are established approaches such as stacking and inverse-error weighting. The authors explicitly state that the contribution is the integration and evaluation rather than a new weighting principle.
  2. Each final surrogate is trained with only one seed, historical evaluation results informed development, and the genuinely new evaluation campaign is limited to Heat I. Thus, the 22-dataset breadth does not fully substitute for independent repetitions or genuinely untouched benchmarks.
  3. Training resources are not matched across methods, and the ensemble requires training many heterogeneous models and their associated coarse models. The paper also acknowledges that net computational savings relative to simulation cost are not measured. It is difficult to judge whether AutoMF is preferable to spending the same compute/data budget on fewer stronger models.
Questions for authors

q1. Can you provide results with multiple independent training seeds, preferably on several representative PDE classes, or evaluate on additional completely untouched datasets? q2. What happens when AutoMF is compared against a compute- and high-fidelity-data-matched alternative—for example, repeated training/ensembling of the strongest single architecture or a comparable fine-only ensemble? q3. The main fitted ensemble minimizes squared relative error, while evaluation uses mean relative L2L_2 error. The appendix shows that directly fitting the reported metric improves the aggregate result by another ~2% Why was the mismatched objective retained as the primary method, and can the authors report direct-relative-L2L_2 fitting on the full 22-dataset benchmark?

Reviewer_wUowOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes AutoMF, an automated ensemble workflow for multifidelity prediction of physical fields. It combines a library of nine surrogates adapted from neural operators, convolutional networks and reduced basis models. The approach aims to reduce errors from uncertain model choice by fitting fixed ensemble weights on a small number of held-out high-fidelity examples. Experiments on 22 datasets spanning seven PDE classes and climate emulation show that AutoMF achieves lower error than selecting a single model on most datasets.

Critical strengths

The paper is well-written and easy to follow.

Although the method itself is a weighted combination of different existing methods, it seems to be practical and data-efficient. With the proposed AutoMF method achieving best and outperforming single sub-model in the mixture regarding experimental results.

The experimental setup spans multiple PDE classes and real-world ERA5 climate emulation, which makes the empirical conclusions comprehensive.

Critical weaknesses

The main contribution of this work is more of engineering than scientific discovery, as the technical novelty is very limited. The weighting rules are standard stacking and inverse-error weighting; and the library members adapt existing architectures. Although the authors acknowledge that the contribution is “integration and evaluation rather than a new weighting principle”, the concern remains.

Each surrogate is trained only with a single seed. Since the gain over the best model in the mixture is relatively small, it is unclear whether the benefit holds across training variability.

The training cost analysis against baselines is not presented, as AutoMF trains nine surrogates and uses extra fine examples, while each baseline is a single model. Adding the cost analysis would provide more insights on model / mixture selection.

Questions for authors

Please refer to the strengths and weaknesses.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.