MENO: Memory-Efficient Neural Operator
Shengyang Xu, Weijun Zhang, Jun Hu, Pengzhan Jin
cs.LG
Quick verdict
MENO: integration-free neural operator, memory independent of resolution; strong accuracy but test-set selection and untuned baselines weaken comparisons.
Summary
MENO is a neural operator for learning PDE solution mappings on arbitrary geometries, built on the Manifold Function Encoder (MFE) previously introduced by the same group. Manifold functions (geometries, coefficients, boundary data) are compressed into a fixed number of Legendre-basis integral tokens, so the global Transformer branch processes n^d tokens regardless of mesh resolution, while a local branch evaluates query points on the output manifold via MFE decoders; global states and their derivatives are injected into local layers and the two predictions are summed. Because training involves no integration over physical points, the authors claim GPU memory is independent of data resolution. On cross-geometry Poisson (9000 samples), MENO reaches 2.07% relative error versus 6.49% for MFE-MIONet; on single-geometry Poisson it marginally beats PCNO (0.669% vs 0.683%) with 1.7% of the parameters, 5.1% of the memory and 8.3% of the training time; on NASA-CRM (105 training shapes) it reports the best pressure/friction errors (6.69%/4.79%); on AhmedML it is second-best behind Transolver-3. Key limitations: checkpoints were selected by highest test accuracy when no validation split existed; baselines are literature-reported numbers, with PCNO hyperparameters transferred untuned; there are no seeds or error bars; the faster-training claim is contradicted on one case (8.40 h vs MFE-MIONet's 4.70 h); and memory-resolution independence is argued structurally but never demonstrated experimentally.
Review complete
Assessment at a glance
5
risk flags identified
● high 1
● medium 3
● low 1
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- MENO's training-stage GPU memory is independent of data resolution because the architecture contains no integration operators, which the authors argue enables scaling to large models.
- MENO accepts arbitrary PDE inputs — arbitrary geometries, arbitrary discretizations, and cross-geometry settings where input functions and output solutions live on different manifolds.
- MENO achieves the best accuracy on most benchmarks tested versus literature-reported results: best on both Poisson cases and NASA-CRM, second-best on AhmedML behind Transolver-3.
- Encoding manifold functions with the MFE (super-algebraically convergent with Legendre/Fourier bases, inherited from [13]) compresses inputs into low-dimensional tokens that decouple model cost from mesh resolution.
Strengths
Structurally grounded efficiency claim
Training contains no integration over physical points: the global branch operates on n^d fixed Legendre-mode tokens (Sec 2-3, Table 2 lists n=8..32), while the local branch is pointwise. Empirically this shows up as 0.239 GB vs PCNO's 4.67 GB on Poisson (Table 4) and 1.94 GB vs PCNO's >80 GB OOM on AhmedML (Table 6).
Cross-geometry capability is demonstrated, not just claimed
Sec 4.1 mixes 2-d manifold functions (domain, coefficient, forcing) with 1-d boundary functions (boundary data) and disk/annulus domains where deformation methods fail; MENO reaches 2.07% error vs 6.49% for MFE-MIONet with 0.721M vs 2.54M parameters (Table 3).
Best accuracy on NASA-CRM among 9 methods
Honest negative and near-miss reporting
Strong reproducibility commitment
Weaknesses
Test-set model selection
Baselines are literature-reported and untuned
Abstract contradicted by its own table
No variance reporting anywhere
Memory-resolution independence never tested
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
MENO: Memory-Efficient Neural Operator — Preliminary Review
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is MENO and how does it differ from FNO or Transolver?
MENO is a neural operator for PDEs that encodes geometries and input functions into a fixed number of Legendre-basis integral tokens via the Manifold Function Encoder. Its training involves no integration over physical mesh points, so GPU memory does not grow with data resolution, whereas FNO/Transolver-style architectures perform integration or attention over all mesh points in every layer, driving memory and cost up with resolution.
What results does MENO achieve in the paper?
On the paper's benchmarks: 2.07% relative error on a cross-geometry Poisson problem versus 6.49% for MFE-MIONet; 0.669% on a fixed-geometry Poisson versus 0.683% for PCNO, using 1.7% of the parameters, 5.1% of the memory and 8.3% of the training time; the best pressure/friction errors on NASA-CRM (6.69%/4.79%); and second-best on AhmedML behind Transolver-3 (3.37%/4.88% versus 2.96%/4.81%).
Is the MENO code available?
Yes. The paper points to https://github.com/jpzxshi/MENO and states that every numerical example can be reproduced with a single command. The benchmarks use public datasets (NASA-CRM and AhmedML CFD benchmarks) plus Poisson datasets from the earlier MFE-MIONet paper.
What are the main caveats about MENO's comparisons?
For benchmarks without validation splits, the checkpoint with the best test accuracy was selected; baseline numbers are copied from other papers and PCNO's hyperparameters were transferred untuned; results are single runs without error bars; and the claimed independence of memory from data resolution is argued structurally but not demonstrated with a resolution-sweep experiment.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 3-5 were extracted only as colorbar tick labels; visual prediction quality could not be assessed.
- The GitHub repository and its single-command reproduction claim were not verified in this review.
- Baseline rows in Tables 5-6 are cited from the Transolver-3 paper and could not be cross-checked against primary sources.
- This is an arXiv preprint (2609.27739); no peer-review verdict exists yet.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.