CoChat

Efficient neural network models of chemical kinetics using a latent asinh rate transformation

Open access
Döppel, Felix A., Votsmeier, Martin · 2023 · Reaction Chemistry & Engineering

Published version of a preprint

The DOI matches its Crossref record; 56 of the 56 references it cites were found in a registry or an index; 3 of the 65 sources its claims cite were checked for support.

Abstract

The proposed latent transformation approach allows building lightweight neural networks that accelerate reactor simulations significantly.

Overview

CoChat Summary

TL;DR

The paper proposes a latent inverse-hyperbolic-sine (asinh) transformation inside neural networks to model steady-state chemical kinetics when rates or source terms change sign. Across two detailed surface mechanisms, the approach improves accuracy and yields reactor-simulation speed-ups of up to 100,000×.

Key findings

  • For preferential oxidation of CO, direct neural-network modeling of sign-changing CO source terms gave relative prediction errors around 100% or higher, conventional asinh preprocessing reduced errors, and the latent asinh version produced application-ready predictions with relative errors below 1%.
  • In a plug-flow reactor simulation for CO preferential oxidation, the neural-network surrogate was reported to be about 50,000× faster than exact steady-state kinetics on the tested hardware, and up to 100,000× faster using a consumer-grade GPU.
  • For ammonia oxidation in the Ostwald process, direct modeling again gave errors near 100%; conventional asinh preprocessing reduced errors to about 15%, while the latent asinh model achieved errors near 1%.
  • The paper reports small models: e.g. CO source terms were modeled with either 34 nodes in 5 hidden layers for CO and 94 nodes in 1 hidden layer for O₂, and the ammonia-oxidation models used 63 nodes in a single hidden layer (≈500 parameters) for each key species.

Methods

  • Two detailed surface-kinetics test cases were used: preferential oxidation of CO on Pt and ammonia oxidation in the Ostwald process; both were simulated under reactor-relevant operating ranges defined in Tables 1 and 2.
  • The dataset comprised 35,000 input-output pairs for each test case, split into 25,000 training, 5,000 validation, and 5,000 test points; neural networks were fully connected, tanh-activated, and trained in PyTorch with full-batch LBFGS.
  • The main comparison was between three strategies: direct modeling, conventional preprocessing with asinh transformation, and the proposed latent asinh transformation, evaluated with mean absolute relative error (MARE) and validated in plug-flow reactor simulations and selectivity calculations.

Limitations

  • The paper states that its previous, reaction-path-analysis-based method was more accurate, but it requires mechanism-specific insight; the new latent asinh method is less accurate than that prior method when such insight is available.
  • For the ammonia-oxidation mechanism, mass transfer was not considered in the validation at Ostwald-process conditions; this is stated in the text.
  • Evident from the text: the study is limited to two detailed mechanistic case studies and steady-state source-term prediction, so broader claims about other mechanisms or transient kinetics are not directly tested here.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1039/d3re00212h

Cites

56 of 56 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Image 1
CC-BY · From the PDF
Image 2
CC-BY · From the PDF
Fig. 1 Plot of the asinh ( x ) function which approximates the logarithm of 2 x for large positive and negative inputs while being linear near the origin.
Fig. 1 Plot of the asinh ( x ) function which approximates the logarithm of 2 x for large positive and negative inputs while being linear near the origin.CC-BY · From the PDF
Image 4
CC-BY · From the PDF
Fig. 2 Comparing conventional and latent training strategy. They differ in how data transformation is applied and which error metric is optimized during training. In both cases, a machine learning model predicts chemical rates using reaction conditions as inputs. The loss is computed to evaluate the prediction accuracy and the model parameters are updated accordingly. Conventionally, the transformation is applied to the data in a preprocessing step. The transformed values asinh ( r ) are then learned by a standard neural network. The disadvantage of this approach is that during training a loss function with respect to the transformed values has to be used instead of the actual error measure of interest. We propose to work with latent (hidden) representations of the transformed data. This means that a model with standard fully connected layers learns a latent representation of the transformed rates. Afterwards, the inverse of the transformation function is applied as a custom output activation in the final layer so that outputs represent the original rates. Hence, the error metric of interest can be optimized during training. If all parameters of the transformation function are fixed before the optimization, the inverse transformation can alternatively be implemented in a customized loss function.
Fig. 2 Comparing conventional and latent training strategy. They differ in how data transformation is applied and which error metric is optimized during training. In both cases, a machine learning model predicts chemical rates using reaction conditions as inputs. The loss is computed to evaluate the prediction accuracy and the model parameters are updated accordingly. Conventionally, the transformation is applied to the data in a preprocessing step. The transformed values asinh ( r ) are then learned by a standard neural network. The disadvantage of this approach is that during training a loss function with respect to the transformed values has to be used instead of the actual error measure of interest. We propose to work with latent (hidden) representations of the transformed data. This means that a model with standard fully connected layers learns a latent representation of the transformed rates. Afterwards, the inverse of the transformation function is applied as a custom output activation in the final layer so that outputs represent the original rates. Hence, the error metric of interest can be optimized during training. If all parameters of the transformation function are fixed before the optimization, the inverse transformation can alternatively be implemented in a customized loss function.CC-BY · From the PDF
Fig. 3 Scheme of the recommended architecture. The neural network takes reaction conditions in form of temperature and partial pressures as input. Those values are transformed and linearly scaled before being fed into conventional hidden layers with tanh activation. The last layer holds a single node per gas phase species and contains y , a latent representation of the transformed target values. A hyperbolic sine activation is applied to obtain outputs in the form of source terms.
Fig. 3 Scheme of the recommended architecture. The neural network takes reaction conditions in form of temperature and partial pressures as input. Those values are transformed and linearly scaled before being fed into conventional hidden layers with tanh activation. The last layer holds a single node per gas phase species and contains y , a latent representation of the transformed target values. A hyperbolic sine activation is applied to obtain outputs in the form of source terms.CC-BY · From the PDF
Image 7
CC-BY · From the PDF
Image 8
CC-BY · From the PDF
Fig. 5 Comparing the CO source term prediction accuracy of different neural network training strategies on 5000 unseen test data randomly sampled from the input range of table 1. Using a standard feed-forward neural network without data transformation ('direct' modeling) does not yield accurate results. Using the asinh transformation conventionally, i.e. in a preprocessing step, reduces the prediction errors considerably. When the asinh transformation is implemented in a latent fashion, the models are even more accurate and yield application ready predictions with relative errors below 1 %. All models contain about 5000 parameters distributed over one (625 nodes), three (48 nodes each) or five hidden layers (34 nodes each) and were trained with 25 000 data points.
Fig. 5 Comparing the CO source term prediction accuracy of different neural network training strategies on 5000 unseen test data randomly sampled from the input range of table 1. Using a standard feed-forward neural network without data transformation ('direct' modeling) does not yield accurate results. Using the asinh transformation conventionally, i.e. in a preprocessing step, reduces the prediction errors considerably. When the asinh transformation is implemented in a latent fashion, the models are even more accurate and yield application ready predictions with relative errors below 1 %. All models contain about 5000 parameters distributed over one (625 nodes), three (48 nodes each) or five hidden layers (34 nodes each) and were trained with 25 000 data points.CC-BY · From the PDF
Fig. 6 Relative prediction error of CO source terms dependent of the total number of learnable parameters in a neural network using the latent hyperbolic sine transformation.
Fig. 6 Relative prediction error of CO source terms dependent of the total number of learnable parameters in a neural network using the latent hyperbolic sine transformation.CC-BY · From the PDF
Fig. 7 Plug-flow reactor model of the preferential oxidation of CO in H 2 rich environments at three different temperatures. The upper part shows CO and O 2 molar fractions along the reactor length. The neural network solution (dotted lines) cannot be visually separated from the exact solution (full lines). The lower part shows the relative difference between both solutions. None of the conditions shown were part of the 25 000 training data which were randomly sampled from the input range of table 1. Feed composition and other details are described in section 2.1.3.
Fig. 7 Plug-flow reactor model of the preferential oxidation of CO in H 2 rich environments at three different temperatures. The upper part shows CO and O 2 molar fractions along the reactor length. The neural network solution (dotted lines) cannot be visually separated from the exact solution (full lines). The lower part shows the relative difference between both solutions. None of the conditions shown were part of the 25 000 training data which were randomly sampled from the input range of table 1. Feed composition and other details are described in section 2.1.3.CC-BY · From the PDF
(a) Comparing the N 2 O source term prediction accuracy of different neural network training strategies on 5000 unseen test data randomly sampled from the input range of table 2. Using a standard feed-forward neural network without data transformation ('direct' modeling) does not yield accurate results. Using the asinh transformation conventionally, i.e. in a preprocessing step, reduces the prediction errors to approximately 15 %. When the asinh transformation is implemented in a latent fashion, the models yield application ready predictions with relative errors near 1 %.
(a) Comparing the N 2 O source term prediction accuracy of different neural network training strategies on 5000 unseen test data randomly sampled from the input range of table 2. Using a standard feed-forward neural network without data transformation ('direct' modeling) does not yield accurate results. Using the asinh transformation conventionally, i.e. in a preprocessing step, reduces the prediction errors to approximately 15 %. When the asinh transformation is implemented in a latent fashion, the models yield application ready predictions with relative errors near 1 %.CC-BY · From the PDF
(b) Models are validated by comparing product selectivities at unseen Ostwald process conditions and zero ammonia conversion (10 % NH 3 in air at 5 bar). The model using asinh data transformation in the conventional way covers the general trend of selectivities but fails at lower temperatures. In contrast, the model using the latent asinh transformation is in excellent agreement with the full model over the whole temperature range. Besides N 2 O, neural network models of NH 3 and N2 are used to fully describe the reaction progress in the system and use the wellknown logarithmic transformation.
(b) Models are validated by comparing product selectivities at unseen Ostwald process conditions and zero ammonia conversion (10 % NH 3 in air at 5 bar). The model using asinh data transformation in the conventional way covers the general trend of selectivities but fails at lower temperatures. In contrast, the model using the latent asinh transformation is in excellent agreement with the full model over the whole temperature range. Besides N 2 O, neural network models of NH 3 and N2 are used to fully describe the reaction progress in the system and use the wellknown logarithmic transformation.CC-BY · From the PDF

Verification

Claims the paper makes

3 Valid · 1 Not valid · 28 Not checked · 80 Own results and interpretations

44 more claims

Create a free account to see all 112 claims and to verify the 10 references still open.

See all claims — free account
  • It may state two separate claims as one. 38
  • Probably not a substantial claim. 36

Create a free account to see these claims.

Citations

Cites

Show all 56 references

Research works printed but not linked: 3

  1. Chase, 1998, J. Phys. Chem. Ref. Data, 14, 927
  2. Klumpers, 2023, SSRN Journal, 1
  3. Chen, 2018, Adv. Neural. Inf. Process. Syst., 31, 1

Cited by

Metadata

Authors
Döppel, F. A., & Votsmeier, M.
Affiliations
Technische Universität Darmstadt (DE)
Corresponding author
Votsmeier, Martin
Venue
Reaction Chemistry & Engineering
Published
2023
Volume
8
Issue
10
Pages
2620-2631
ISSN
2058-9883 (ISSN-L)
OpenAlex
W4383960100
Cited by
24
References
55
Citations by year
2026: 6 · 2025: 8 · 2024: 9 · 2023: 1
FWCI
2.38
Type
journal article
Language
en
Open access
Locations
  • Reaction Chemistry & Engineering (publishedVersion, cc-by)
  • ChemRxiv (acceptedVersion, cc-by-nc-nd)
Funders
Bundesministerium für Bildung und Forschung (05M20RDA)
Topics
  • Physical Sciences › Physics and Astronomy › Statistical and Nonlinear Physics › Model Reduction and Neural Networks
  • Physical Sciences › Computer Science › Artificial Intelligence › Neural Networks and Applications
  • Physical Sciences › Engineering › Control and Systems Engineering › Control Systems and Identification
Keywords
reaction kinetics, neural network surrogate model, lightweight neural networks

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-10-10, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-10 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol and qwen3.8-omni-flash).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.