acceptodds
Withdrawn from ICLR 2027

DNA: A Double-Helix Hierarchical Deep Interaction Network Architecture for Industrial Digital Twin

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

High-fidelity modeling is fundamental to industrial digital twin, yet a trade-off persists between the physical transparency of mechanistic models and the superior predictive power of data-driven approaches. Existing hybrid methods often fail to bridge this gap, suffering from shallow integration and rigid one-shot modeling strategies that struggle with complex industrial systems. To address these issues, we propose a . Inspired by the double-helix structure of biological , our architecture intertwines mechanistic and data-driven modeling strands through two novel interaction strategies, forming a symbiotic architecture in which the two strands mutually enhance predictive capability. Furthermore, the hierarchical structure of DNA allows it to model complex systems in a multi-stage manner through two tailored deep learning frameworks. We evaluate the proposed approach on two distinct industrial systems, where DNA achieves reductions in MAE of 3.704% and 3.328%, respectively, compared to the strongest baseline. These results demonstrate the effectiveness of DNA and suggest its potential to provide new insights for the development of industrial digital twins.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Deep learning, Industrial digital twin, Industrial modeling, Data-driven modeling, Three-phase pipeline, Chemical engineering

Reviews

Among:
Shape:
1234this paper, Bottom 25%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.75 over 4 reviews, above 25% of the others.
Reviewer_HThDOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces DNA, a hierarchical hybrid architecture for industrial digital twin, inspired by the structure of biological DNA. It combines mechanistic features and a neural network in two stages. Stage I uses an attention network (MASA) guided by the mechanistic model’s structure to predict each property. It also generates coefficients for a predefined correction function FF applied to the mechanistic output ymey_{me}. Stage II is a GCN-Transformer that predicts liquid holdup (HL) and outlet pressure (P). The approach is evaluated on a three phase oil-gas-water pipeline, and on acentric factor prediction for petroleum pseudo-components.

Critical strengths

The paper’s most interesting point is that data driven modeling may leave behind an improved mechanistic simulator that operates independently. In safety-critical process industries, the mechanistic model is what gets deployed, so using data to improve that model rather than replace it addresses a real need. The downstream experiment is useful evidence that correcting upstream properties improves this pipeline.

Critical weaknesses
  1. line 250, Eq. (4) shows the coefficients are functions of the input, but Table 7 uses fixed expressions. If the coefficients depend on the input, the network is needed at inference, so the model cannot be deployed independently, contrary to the paper's contribution (1), lines 84–86. If the coefficients are global, any state dependence the network learned during training is discarded at deployment.
  2. The paper does not show that DNA is better than simply fitting a correction curve. With fixed coefficients, FF depends on the input only through ymey_{me}. In this case, if the original model gives the same prediction for two different operating conditions, the corrected model will receive the same correction and give the same result. So the model cannot recover a pressure or composition effect that is missing from the original prediction.
  3. Following point 2, the benchmark does not test this clearly. Under the paper’s assumptions (Appendix C and Table 5), 6 of 8 Stage I tasks vary only in T. Where the original prediction consistently increases or decreases with temperature, fitting a correction to that prediction is effectively fitting another temperature-dependent curve. The results are consistent with this. In Table 1, for liquid viscosity, Cor barely depends on the architecture (Vis.l MAE 43.6-44.8 across 12 networks). DNA’s proposed model gives best results on Cor in only two datasets (out of nine). The two largest gaps between DD and Cor occur on multi-input targets (AF 0.087 vs. 0.279; Rho.l 2.43 vs. 6.58) where the limitation in point 2 exactly applies.
  4. In the ablations, the advantage of the full model is not convincing. Many wins are only ~1%. Especially given that they seem to be single runs, these differences could be within run-to-run noise. On Rho.l, three of four ablations achieve lower DD MAE than the full model.
  5. Hybrid-model baselines are missing. The paper positions DNA against structured hybrid modeling with serial, parallel and combined architectures (lines 52–56; Appendix A, Fig. 4), but none of these is included in the experiments. All 12 baselines are neural backbones run inside the DNA framework, and many of them are time-series forecasters applied to static rows of numbers. A fair comparison needs at least a parallel residual hybrid, for example, yme+NN(X,Z)y_{me} + NN(X, Z), and a serial hybrid network that takes ymey_{me} as an input. Without them, the experiments compare model backbones within DNA, rather than DNA against existing hybrid approaches.
Questions for authors
  1. The correction functions have a predefined form with only a few coefficients (Section 4.5). If these coefficients are fitted directly, e.g. by least-squares regression of the true values on ymey_{me} using the same training data and no neural network, how does the corrected model’s accuracy compare with DNA’s?
  2. How was uncertainty determined? Lines 298–302 describe the targets as measurements, but Cp.w DD MAE is 0.724 J/(kg·K), about 0.02% of its value, which seems below plausible measurement uncertainty.
Reviewer_bnv3OpenReview ↗
Rating
2
Confidence
1
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a new hybrid architecture for combining mechanistic and data-driven approaches for modeling industrial processes for systems that require multiscale modeling.

Industrial digital twins require high-fidelity models to enable monitoring, prediction and optimization and mechanistic models are usually preferred. Mechanistic models themselves often have simplifying assumptions which can introduce systematic biases under unfamiliar conditions and require substantial modeling effort to construct.

Data-driven models hold the promise of alleviating these limitations and have been shown to outperform mechanistic models in terms of prediction accuracy, efficiency and cost. However, their predictions may not always be grounded in physical principles. This has led to hybrid modelng approaches combining mechanistic and data-driven features. Current hybrid approaches are simplistic amounting to stacking of mechanistic and data-driven parts following a unidirectional flow. Moreover they follow a ‘one-shot’ strategy that attempts to predict the output in a single step.

In the paper’s proposal the two modules are joined in a DNA-like sequence and the interaction resembles the base-pair structure while allowing bidirectional interaction between the two modules. In the proposed architecture data-driven models embed enhancements into mechanistic models which allows them to be independently deployed.

The method starts by building an original mechanistic model for a 3 phase pipeline process. The stage I model uses the original operating conditions to calculate physical properties of different phases. The stage II models takes the input and output and computes another output for downstream variables. The mathematical difference is that stage I is an explicit function of variables and stage II represents an implicit constraint.

Given the mechanistic model the stage I data-driven model takes data from the mechanistic model to produce outputs and a correction which are fed back into the mechanistic model to produce corrected output. The state II data model incorporates data from the stage II mechanistic model with graph networks and produces predictions.

Critical strengths

The hybrid method proposes combining data-driven and mechanistic modeling in a two-way connection, also improving the mechanistic model. The paper compares against a number of baselines including transformer and graph methods and performs better than the baselines.

Critical weaknesses

The main difficulty with the paper is that the description remains very abstract. There is little to no insight into the structure of the data or the task (or process) for the reader to judge whether the proposed components and design are appropriate for that purpose.

It is not clear to me from the paper where the graph structure in the mechanistic model comes from or how it is constructed. It appears that mechanistic function composition is treated as a graph, but it is unclear how and why it would be an appropriate design.

The paper doesn’t explain what complex topologies are present for which graph networks would be inappropriate (line 202)

The function of the correction function is unclear.

One of the main claims of the paper: improving the mechanistic model so that it could be used standalone is unclear.

Questions for authors

What is the form of the bidirectional mechanism? especially how does feedback from the data-driven component to the mechanistic component work?

What are intermediate results Z of the mechanistic model in line 146

What is the correction function in linear 149 produced by the stage ! data-driven model? how does it refine predictions of the mechanistic model?

Is data-driven modeling of intermediate variables also causal? From the description on page 5 (top) it seems that direction quantities are said to depend on some quantities and independent of others. Is this causal or statistical dependence/independence? Does this mean the digital twin also enforces causal relations for intermediate variables? If they are causal, are there any advantages to model also the correlations? could prediction improve if other correlations are also modeled?

In line 257-259 the modeling task is to solve the system of equations? is that what is meant by predicting unknown variables? or is something else meant here?

Formatting etc.

The citations have missing parentheses. Table 1 is not very readable.

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

The authors propose in a quite ad-hoc and superficial manner a framework called Double-Helix Hierarchical Deep Interaction Network Architecture (DNA) and they claim it is usable for industrial digital twin. Unfortunately, this model is only tested on a very particular simulation setting involving a three-phase pipeline and a very particular model. To make the claim that DNA is valid and applicable at industrial scale (industrial digital twin) is too much. The authors fail to justify the properties of this model. With the whole model construction presented in a very ad-hoc way, without any justification about what type of dynamics can be captured and how well these dynamics can be captured, this construction very very limited value. Thus, I recommend authors to provide rigorous analytical justifications of the merits and features of the proposed DNA modeling framework.

Critical strengths

The DNA model construction presents some interesting combination of ideas. However, the whole construction of DNA modeling framework is lacking theoretical justification. No analytical justification for the resulting modeling capabilities is presented. No justifications for the modeling accuracy and global convergence are presented. Overall, a superficial paper.

Critical weaknesses

As mentioned before, the authors need to 1) justify rigorously the modeling merits of the proposed DNA framework? What are its global modeling capabilities? What types of problems is fit for? What is it accuracy? How accurate general dynamics can be captured? What about its global convergence? 2) the set of baseline models is superficial. 3) why this model can be considered as a industrial digital twin for any application?

This model: "the model can simulate the behavior of the physical system and serve as its digital twin. This task can be formulated as: min M EX∼X [|M(X) − O(X)|], where E[·] denotes the mathematical expectation and X denotes the distribution of operating conditions." is not good in the majority of practical problems. You are just considering the mean of the error. Not sufficient.

Questions for authors

The simulation experiments are superficial. Claims such as "Extensive experiments on two distinct industrial systems demonstrate effectiveness and robustness of our DNA architecture" should not be made in a professional paper. Please justify analytically this claim.

You claim "In particular, DNA is an industrial modeling architecture rather than merely a specific model." Please provide a solid and clear justification for this claim.

Please justify this claim "DNA provides a general architecture for integrating mechanistic and data-driven models across multiple modeling stages."

The very ad-hoc constructions in Section 3 Methodology need to be justified for the standpoint of algorithm's accuracy, modeling capability, convergence, robustness, generalization abilities etc.

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

DNA is a hybrid architecture for industrial digital twins. Mechanistic and data-driven model chains interact "bidirectionally" across a hierarchy of stages, with a mechanism-aware attention module (MASA) that restricts each variable's attention to the variables it physically depends on. It's evaluated on a three-phase pipeline and on acentric-factor prediction. Pairing ML with mechanistic models is worthwhile, but the paper doesn't show that this particular design matters.

Critical strengths

None

Critical weaknesses

The baselines are the wrong ones. The comparison is against generic regressors and time-series models (iTransformer, TimesNet, TimeXer, UMixer, etc.). There are no hybrid or physics-informed baselines: no residual/gray-box hybrids, no PINNs, and no structured hybrid models, even though the paper itself frames those as the prior work it improves on. As set up, the experiments can't support the claim that DNA beats existing hybrid approaches.

The "double helix" is mostly a metaphor. As far as I can tell, the mechanistic side is corrected by a learned correction function, and the data-driven side takes mechanistic inputs. That's a fairly standard residual hybrid. The paper needs a concrete ablation showing that the bidirectional, multi-stage coupling beats simple residual correction with the same backbone.

The "Cor" improvements over MM aren't a fair win. Any reasonable learned correction beats a biased mechanistic model, and the paper itself shows nearly every baseline does so inside the DNA framework. That says the regressor helps, not that the architecture does.

The data look simulated and narrow. There are only two systems, and the pipeline data (15,180 samples) appear to be generated rather than measured. Nothing shows behavior in the extrapolation and distribution-shift regimes where the paper says mechanistic grounding is supposed to help.

The efficiency claim is backwards. DNA has the fewest FLOPs but is 2–3× slower at inference than most baselines (32.9 ms vs about 12 ms), because MASA runs sequentially. The paper calls DNA "lightweight", but it's slower in practice.

Questions for authors

The numbers don't match. The abstract says 3.704% / 3.328%, while Sec. 4.3 says MAE drops by 7.80% on average against MultiPatchFormer. Which is it, and against what?

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

Related papers

Discussion (0)

Sign in to comment.