Vortex-Induced Drag Forecast for Cylinder in Non-uniform Inflow

Dixia Fan; Haoyang Hu; Huimin Wang; Jiashun Guan; Tianfang Hao; Yunxiao Ren

arxiv: 2506.21075 · v1 · submitted 2025-06-26 · ⚛️ physics.flu-dyn

Vortex-Induced Drag Forecast for Cylinder in Non-uniform Inflow

Jiashun Guan , Haoyang Hu , Tianfang Hao , Huimin Wang , Yunxiao Ren , Dixia Fan This is my paper

Pith reviewed 2026-05-19 08:09 UTC · model grok-4.3

classification ⚛️ physics.flu-dyn

keywords vortex-induced dragcircular cylindernon-uniform inflowneural networkdirect numerical simulationpressure sensorsflow separationdrag forecasting

0 comments

The pith

A modified neural network using upstream velocity and optimized pressure signals forecasts vortex-induced drag fluctuations on a cylinder with R^2 of 0.75.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper develops a physics-based data-driven approach to forecast drag forces from vortices on a circular cylinder when the incoming flow varies spatially, a common challenge in engineering at moderate Reynolds numbers. Traditional pressure-only models struggle with the coupled vortex dynamics, so the authors combine pressure signals from the cylinder surface with velocity measurements taken upstream to calibrate the inflow conditions. Using direct numerical simulations at Reynolds number 4000, they optimize sensor locations and velocity components through iterative training of a fully connected neural network. The resulting model reaches an R^2 score of 0.75 when predicting large drag coefficient swings between 0.2 and 1.2 over the next time unit. The best-performing sensor positions align with regions where flow separates, which the authors link directly to the generation of vortex-induced drag.

Core claim

A modified fully connected neural network integrates upstream velocity measurements for inflow calibration with pressure-signal inputs and, after optimization on DNS data at Re=4000, achieves an R^2 score of 0.75 in forecasting high-amplitude drag coefficient fluctuations (Cd from 0.2 to 1.2) within a future time window of one time unit. Model performance exhibits an exponential scaling with the number of optimized pressure inputs, and the selected sensor placements correspond to the physical mechanism in which flow separation dynamics govern vortex-induced drag generation.

What carries the argument

Modified fully connected neural network that adds upstream velocity measurements as inflow calibration to pressure-signal inputs, with iterative optimization of sensor placements.

If this is right

An exponential scaling relates model performance to the number of optimized pressure signal inputs.
Sparsely distributed but optimized pressure sensors can still deliver useful predictive skill.
Sensor placements that match flow separation locations capture the main mechanism of vortex-induced drag.
The strategy provides a scalable route to forecasting statistics in turbulent flows under engineering conditions.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same optimization procedure could be repeated on experimental rather than simulated data to test robustness outside controlled DNS conditions.
Sparse sensor sets tuned this way might lower the instrumentation cost for monitoring vortex-induced loads on offshore or wind structures.
The inflow-calibration idea may extend to other bluff-body problems where upstream disturbances affect wake dynamics.

Load-bearing premise

The neural network trained and optimized on DNS data at Re=4000 with specific non-uniform inflow setups will generalize to predict drag accurately under other real-world non-uniform inflow variations or different Reynolds numbers.

What would settle it

Apply the optimized model to new DNS or experimental data at a different Reynolds number such as Re=1000 or with a qualitatively different inflow profile and check whether the R^2 for high-amplitude drag forecasts falls substantially below 0.75.

read the original abstract

In this letter, a physics-based data-driven strategy is developed to predict vortex-induced drag on a circular cylinder under non-uniform inflow conditions - a prevalent issue for engineering applications at moderate Reynolds numbers. Traditional pressure-signal-based models exhibit limitations due to complex vortex dynamics coupled with non-uniform inflow. To address this issue, a modified fully connected neural network (FCNN) architecture is established that integrates upstream velocity measurements (serving as an inflow calibration) with pressure-signal-based inputs to enhance predictive capability (R^2 ~ 0 to 0.75). Direct numerical simulations (DNS) at Reynolds number Re = 4000 are implemented for model training and validation. Iterative optimizations are conducted to derive optimized input configurations of pressure sensor placements and velocity components at upstream locations. The optimized model achieves an R^2 score of 0.75 in forecasting high-amplitude drag coefficient fluctuations (C_d=0.2 - 1.2) within a future time window of one time unit. An exponential scaling between model performance and optimized pressure signal inputs is observed, and the predictive capability of sparsely distributed but optimized sensors is interpreted by the scaling. The optimized sensor placements correspond to the physical mechanism that the flow separation dynamics play a governing role in vortex-induced drag generation. This work advances machine learning applications in fluid-structure interaction systems, offering a scalable strategy for forecasting statistics in turbulent flows under real-world engineering conditions.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper gets R^2 up to 0.75 for short-term drag forecasting on a cylinder at Re=4000 by adding upstream velocity inputs and optimizing pressure sensor locations, but everything stays inside one narrow DNS dataset.

read the letter

The core result is straightforward: a modified FCNN that folds in upstream velocity components as inflow calibration, plus iterative optimization of which pressure sensors to use, lifts forecasting performance on high-amplitude Cd fluctuations from roughly zero to 0.75 over a one-time-unit window. They also report exponential scaling of accuracy with the number of optimized sensors and map the best placements back to flow-separation physics. That combination and the scaling observation are the actual new pieces relative to earlier pressure-only models. The work is honest about its engineering focus and gives a usable benchmark for this specific setup. The limitation is equally clear. All training, optimization, and validation sit on DNS at fixed Re=4000 with the particular non-uniform inflows they chose. No runs at other Reynolds numbers, no fresh inflow profiles, and no held-out test cases that would show whether the learned mapping travels. The abstract gives no error bars, no cross-validation protocol, and no mention of how sensitive the R^2 is to the optimization details. Because the sensor and velocity selections are tuned on the same data used to score the model, the reported number is a data-driven fit rather than an independent prediction. Readers who need a quick, practical tool for moderate-Re cylinder drag under controlled non-uniform conditions will find the method and the scaling plot useful. Anyone who wants evidence that the approach generalizes to real-world variations will have to wait for more tests. The paper is coherent on its own terms and shows clear thinking about the inputs, so it is worth sending to referees who can ask for those generalization checks.

Referee Report

3 major / 2 minor

Summary. The manuscript develops a physics-based data-driven strategy using a modified fully connected neural network (FCNN) that combines upstream velocity measurements with pressure signals to forecast vortex-induced drag on a circular cylinder in non-uniform inflow. Direct numerical simulations at Re=4000 are used for training and validation; iterative optimization selects pressure-sensor locations and upstream velocity components, yielding an R² improvement from ~0 to 0.75 for high-amplitude Cd fluctuations (0.2–1.2) over a one-time-unit forecast horizon. The work reports exponential scaling of performance with the number of optimized sensors and interprets the selected placements as corresponding to flow-separation dynamics.

Significance. If the mapping generalizes, the approach could supply a practical, sensor-optimized route to drag forecasting in engineering flows where non-uniform inflow is common. The explicit optimization loop together with the post-hoc physical interpretation of sensor locations is a constructive element that goes beyond pure black-box regression. At present, however, all quantitative results are obtained on a single Reynolds number and a narrow family of inflow realizations, so the engineering claim remains provisional.

major comments (3)

[Results on model optimization and performance evaluation] The optimization of pressure-sensor placements and upstream velocity components (described in the iterative-optimization procedure) is performed on the identical DNS dataset used for both training and reported validation. Consequently the quoted R²=0.75 is an in-sample fit rather than an independent prediction; no held-out inflow profiles or Reynolds-number variations are shown to test whether the learned mapping survives changes in the inflow distribution.
[Abstract] The abstract asserts that the method offers “a scalable strategy for forecasting statistics in turbulent flows under real-world engineering conditions,” yet the only quantitative evidence is confined to Re=4000 with the specific non-uniform inflow realizations employed in the DNS campaign. Absence of any cross-Re or cross-inflow test directly undermines the central engineering claim.
[Results on model optimization and performance evaluation] No error bars, cross-validation protocol, or statistical significance assessment accompany the reported R² values (including the transition from ~0 to 0.75). Without these, it is impossible to judge whether the improvement is robust or sensitive to the particular train/validation split.

minor comments (2)

[Abstract] The phrase “C_d=0.2 - 1.2” should clarify whether this interval refers to absolute coefficient values or to fluctuation amplitude about the mean.
[Results on model optimization and performance evaluation] The exponential scaling between model performance and number of optimized sensors is stated but not accompanied by a figure or table showing the raw data points; a supplementary plot would strengthen the claim.

Simulated Author's Rebuttal

3 responses · 1 unresolved

We thank the referee for the constructive and detailed comments on our manuscript. These observations highlight important aspects of validation and scope that we will address in the revision. We respond to each major comment below.

read point-by-point responses

Referee: [Results on model optimization and performance evaluation] The optimization of pressure-sensor placements and upstream velocity components (described in the iterative-optimization procedure) is performed on the identical DNS dataset used for both training and reported validation. Consequently the quoted R²=0.75 is an in-sample fit rather than an independent prediction; no held-out inflow profiles or Reynolds-number variations are shown to test whether the learned mapping survives changes in the inflow distribution.

Authors: We agree that performing the sensor optimization and model fitting on the full dataset renders the reported R² an in-sample result. In the revised manuscript we will adopt a strict train/test separation: the iterative optimization loop and network training will be executed exclusively on a training subset of the DNS realizations, while all performance metrics (including the R² improvement) will be evaluated on a held-out test set containing distinct inflow profiles. This change will be documented with explicit descriptions of the split ratios and the inflow characteristics of the test cases. With respect to Reynolds-number variations, the present study is confined to Re = 4000; demonstrating cross-Re generalization would require new DNS campaigns at additional Reynolds numbers and lies outside the scope of the current work. revision: partial
Referee: [Abstract] The abstract asserts that the method offers “a scalable strategy for forecasting statistics in turbulent flows under real-world engineering conditions,” yet the only quantitative evidence is confined to Re=4000 with the specific non-uniform inflow realizations employed in the DNS campaign. Absence of any cross-Re or cross-inflow test directly undermines the central engineering claim.

Authors: We accept that the abstract phrasing is broader than the quantitative evidence supplied. We will revise the abstract to state that the approach is demonstrated at Re = 4000 for the family of non-uniform inflows examined in the DNS, and that it offers a promising route toward scalable forecasting in engineering flows, subject to further validation across regimes. The revised wording will align the claim with the concrete results while preserving the forward-looking interpretation already present in the discussion section. revision: yes
Referee: [Results on model optimization and performance evaluation] No error bars, cross-validation protocol, or statistical significance assessment accompany the reported R² values (including the transition from ~0 to 0.75). Without these, it is impossible to judge whether the improvement is robust or sensitive to the particular train/validation split.

Authors: We acknowledge the absence of statistical quantification. In the revision we will introduce a cross-validation protocol (multiple random train/test partitions or k-fold cross-validation) and will report the mean R² together with standard deviations or error bars for the key performance figures. This will allow readers to assess the sensitivity of the reported improvement to the data split. revision: yes

standing simulated objections not resolved

Demonstration of performance at Reynolds numbers other than 4000, which would require additional direct numerical simulations beyond the resources of the present study.

Circularity Check

1 steps flagged

R²=0.75 'forecast' obtained by iterative optimization of sensor/velocity inputs on the same Re=4000 DNS used for training and validation

specific steps

fitted input called prediction [Abstract]
"Direct numerical simulations (DNS) at Reynolds number Re = 4000 are implemented for model training and validation. Iterative optimizations are conducted to derive optimized input configurations of pressure sensor placements and velocity components at upstream locations. The optimized model achieves an R^2 score of 0.75 in forecasting high-amplitude drag coefficient fluctuations (C_d=0.2 - 1.2) within a future time window of one time unit."

Training, validation, and iterative selection of both sensor placements and velocity inputs all occur on the same DNS realizations at Re=4000. The reported R² is therefore the performance of a model whose inputs were chosen to maximize that same metric on the identical data distribution, making the 'forecast' statistically forced rather than an out-of-distribution prediction.

full rationale

The paper's central performance claim is an optimized FCNN that reaches R²=0.75 on high-amplitude Cd fluctuations. This result is produced by training the network and iteratively selecting pressure-sensor locations plus upstream velocity components on the identical DNS dataset at fixed Re=4000. No held-out test set from different Re or inflow profiles is reported, so the quoted forecasting metric reduces to a data-driven fit rather than an independent prediction. The exponential scaling and physical interpretation of placements are likewise post-hoc on the same narrow distribution. This matches the 'fitted input called prediction' pattern with partial circularity; the work remains useful as an empirical demonstration but does not constitute a first-principles derivation.

Axiom & Free-Parameter Ledger

2 free parameters · 1 axioms · 0 invented entities

The central claim rests on data-driven fitting of a neural network to DNS-generated pressure and velocity signals plus post-hoc optimization of sensor inputs; no new physical axioms or entities are introduced beyond standard incompressible flow assumptions.

free parameters (2)

Neural network weights and biases
Fitted during supervised training on DNS data to map inputs to future drag values.
Optimized pressure sensor placements and upstream velocity components
Selected via iterative optimization to maximize R² on the training/validation data.

axioms (1)

domain assumption Incompressible Navier-Stokes equations govern the flow at Re=4000
Underlying assumption for generating the DNS training data.

pith-pipeline@v0.9.0 · 5795 in / 1315 out tokens · 46757 ms · 2026-05-19T08:09:42.718819+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

modified fully connected neural network (FCNN) architecture … integrates upstream velocity measurements … with pressure-signal-based inputs … R² ∼ 0 → 0.75 … exponential scaling between model performance and optimized pressure signal inputs
IndisputableMonolith/Foundation/DimensionForcing.lean alexander_duality_circle_linking unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

optimized sensor placements correspond to the physical mechanism that the flow separation dynamics play a governing role

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.