# KPPax: preprocess differentiable kinetic ODE for machine learning in atmospheric chemistry [Canonical reading](https://escapemeridian.com/readings/kppax-preprocess-differentiable-kinetic-ode-for-machine-550b604f) Wenqing Peng, Michael Boy and Zhi-Song Liu present KPPax, which turns atmospheric chemistry mechanisms written in the KPP language into differentiable, just-in-time compiled and vectorisable chemistry code in JAX. This is intended to make existing mechanisms usable with machine-learning tools for parameter estimation and hybrid modelling. The authors report two demonstrations using the POLLU benchmark mechanism. In one, the Fisher information matrix predicts which rate coefficients a flow-tube-like design can identify; noisy synthetic estimation supports the predicted uncertainties for nine identifiable coefficients, while showing that some fast reactions are constrained only by their ratios. In the other, a neural-network correction trained on retrievals and then fine-tuned through the chemical model reduces the error of a biogenic correction by a factor of 2.4–5 and recovers information on low-temperature emissions that individual retrievals lose, according to the authors. These results suggest a route to combining chemical models with machine learning, but the abstract does not establish performance beyond the stated benchmark and demonstrations. How well does KPPax perform with other atmospheric chemistry mechanisms and real-world observations? Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims. ## Original source arXiv cs.AI · Article Original source date: Oct 27, 2026 (2026-10-27). Date basis: Date basis not provided. Added to Escape Meridian: 2026-11-14T05:17:19.067Z. First observed: 2026-10-28T08:06:36.181Z. Retrieved: 2026-10-30T00:34:07.514Z. - [KPPax: preprocess differentiable kinetic ODE for machine learning in atmospheric chemistry](https://arxiv.org/abs/2610.37655) ## Limits The abstract describes two applications using the POLLU benchmark mechanism; it does not report results across other mechanisms. The parameter-estimation uncertainties are checked with noisy synthetic estimation, and the emission-correction result is described for the stated application; the abstract does not report validation with real-world observations. The reported reduction in error and recovery of information concern the authors' demonstration, not a general guarantee of performance. ## Open question How well does KPPax perform with other atmospheric chemistry mechanisms and real-world observations? ## Source-stated KPPax generates differentiable, just-in-time compiled, vectorisable chemistry code in JAX from mechanisms written in KPP. > KPPax, a kinetic preprocessor that generates differentiable, just-in-time compiled, and vectorisable chemistry code in JAX from mechanisms written in the KPP language - [Supporting source](https://arxiv.org/abs/2610.37655) ## Source-stated The authors say KPPax uses sparse operations in log space and diffrax solvers and adjoints to obtain sensitivities without hand-derived adjoint code. > integrates the stiff system with the implicit solvers and the forward- and reverse-mode adjoints of the \\texttt{diffrax} library, so that exact sensitivities are obtained without hand-derived adjoint code - [Supporting source](https://arxiv.org/abs/2610.37655) ## Source-stated In a POLLU benchmark demonstration, the Fisher information matrix predicts which of 25 rate coefficients a flow-tube-like design can identify; noisy synthetic estimation confirms predicted uncertainties for nine identifiable coefficients to within about 30%. > noisy synthetic estimation confirms the predicted uncertainties of the nine identifiable rate coefficients to within about 30\\% - [Supporting source](https://arxiv.org/abs/2610.37655) ## Source-stated The authors report that a neural-network correction reduces the error of the biogenic correction by a factor of 2.4–5 in their second POLLU benchmark application. > which reduces the error of the biogenic correction by a factor of 2.4--5 - [Supporting source](https://arxiv.org/abs/2610.37655) ## Escape Meridian interpretation The abstract reports demonstrations using the POLLU benchmark mechanism, not results across other mechanisms or real-world observations. > We demonstrate two applications with the POLLU benchmark mechanism. - [Supporting source](https://arxiv.org/abs/2610.37655) ## Source and contributor context Peng, Wenqing, Boy, Michael, Liu, Zhi-Song · Article author These are credits stated by the source; they do not establish a profile match or authorship of every passage. ## Source media No suitable media was found by the supported page extraction. ## Terms and references Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement. ## Machine learning Methods that let computers learn patterns from examples instead of following rules written by hand. - [Definition reference](https://developers.google.com/machine-learning/glossary#machine-learning) ## Benchmark A defined set of tests used to compare systems. Results depend on the tasks and measures selected. - [Definition reference](https://arxiv.org/abs/2211.09110) ## Neural network A system of connected simple calculations, loosely inspired by the brain, that learns patterns from data. - [Definition reference](https://developers.google.com/machine-learning/glossary#neural-network) ## Fine-tuning Further training of an existing model to adapt it to a particular task or use. - [Definition reference](https://developers.google.com/machine-learning/glossary#fine-tuning) ## arXiv arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed. - [Explore this context](/entities/arxiv) - [Reference](https://arxiv.org/) - [Reference](https://info.arxiv.org/help/api/index.html) ## TIME TIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services. - [Explore this context](/entities/time) - [Reference](https://time.com/) ## Editorial provenance Version 1. Publication ID: e59bdeab-00cc-46cb-8b5b-c55cb001260e. Prepared by gpt-6-luna; grounding review by gpt-6-luna. Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-14T05:17:19.067Z. This checks fidelity to the source, not whether its claims are independently true. Editorial selection: accept; reasons: material\_development. Observation: 65f90f695921d3839caa19ffcd2e66cd59eb935af766f0196957c06049d361ec. Source text hash: 03e2ece0ce6d7240a06549fb2e4d99a4fedfe4b461e97772f9a4458edd3cf926. ## Revision and source-check history ### Version 1 — 2026-11-14T05:17:19.067Z Wenqing Peng, Michael Boy and Zhi-Song Liu present KPPax, which turns atmospheric chemistry mechanisms written in the KPP language into differentiable, just-in-time compiled and vectorisable chemistry code in JAX. This is intended to make existing mechanisms usable with machine-learning tools for parameter estimation and hybrid modelling. The authors report two demonstrations using the POLLU benchmark mechanism. In one, the Fisher information matrix predicts which rate coefficients a flow-tube-like design can identify; noisy synthetic estimation supports the predicted uncertainties for nine identifiable coefficients, while showing that some fast reactions are constrained only by their ratios. In the other, a neural-network correction trained on retrievals and then fine-tuned through the chemical model reduces the error of a biogenic correction by a factor of 2.4–5 and recovers information on low-temperature emissions that individual retrievals lose, according to the authors. These results suggest a route to combining chemical models with machine learning, but the abstract does not establish performance beyond the stated benchmark and demonstrations. How well does KPPax perform with other atmospheric chemistry mechanisms and real-world observations? ## Original source arXiv cs.AI · Article Original source date: Oct 27, 2026 (2026-10-27). Date basis: Date basis not provided. Added to Escape Meridian: 2026-11-14T05:17:19.067Z. First observed: 2026-10-28T08:06:36.181Z. Retrieved: 2026-10-30T00:34:07.514Z. - [KPPax: preprocess differentiable kinetic ODE for machine learning in atmospheric chemistry](https://arxiv.org/abs/2610.37655) ## Limits The abstract describes two applications using the POLLU benchmark mechanism; it does not report results across other mechanisms. The parameter-estimation uncertainties are checked with noisy synthetic estimation, and the emission-correction result is described for the stated application; the abstract does not report validation with real-world observations. The reported reduction in error and recovery of information concern the authors' demonstration, not a general guarantee of performance. ## Open question How well does KPPax perform with other atmospheric chemistry mechanisms and real-world observations? ## Source-stated KPPax generates differentiable, just-in-time compiled, vectorisable chemistry code in JAX from mechanisms written in KPP. > KPPax, a kinetic preprocessor that generates differentiable, just-in-time compiled, and vectorisable chemistry code in JAX from mechanisms written in the KPP language - [Supporting source](https://arxiv.org/abs/2610.37655) ## Source-stated The authors say KPPax uses sparse operations in log space and diffrax solvers and adjoints to obtain sensitivities without hand-derived adjoint code. > integrates the stiff system with the implicit solvers and the forward- and reverse-mode adjoints of the \\texttt{diffrax} library, so that exact sensitivities are obtained without hand-derived adjoint code - [Supporting source](https://arxiv.org/abs/2610.37655) ## Source-stated In a POLLU benchmark demonstration, the Fisher information matrix predicts which of 25 rate coefficients a flow-tube-like design can identify; noisy synthetic estimation confirms predicted uncertainties for nine identifiable coefficients to within about 30%. > noisy synthetic estimation confirms the predicted uncertainties of the nine identifiable rate coefficients to within about 30\\% - [Supporting source](https://arxiv.org/abs/2610.37655) ## Source-stated The authors report that a neural-network correction reduces the error of the biogenic correction by a factor of 2.4–5 in their second POLLU benchmark application. > which reduces the error of the biogenic correction by a factor of 2.4--5 - [Supporting source](https://arxiv.org/abs/2610.37655) ## Escape Meridian interpretation The abstract reports demonstrations using the POLLU benchmark mechanism, not results across other mechanisms or real-world observations. > We demonstrate two applications with the POLLU benchmark mechanism. - [Supporting source](https://arxiv.org/abs/2610.37655) ## Source and contributor context Peng, Wenqing, Boy, Michael, Liu, Zhi-Song · Article author These are credits stated by the source; they do not establish a profile match or authorship of every passage. ## Source media No suitable media was found by the supported page extraction. ## Terms and references Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement. ## Machine learning Methods that let computers learn patterns from examples instead of following rules written by hand. - [Definition reference](https://developers.google.com/machine-learning/glossary#machine-learning) ## Benchmark A defined set of tests used to compare systems. Results depend on the tasks and measures selected. - [Definition reference](https://arxiv.org/abs/2211.09110) ## Neural network A system of connected simple calculations, loosely inspired by the brain, that learns patterns from data. - [Definition reference](https://developers.google.com/machine-learning/glossary#neural-network) ## Fine-tuning Further training of an existing model to adapt it to a particular task or use. - [Definition reference](https://developers.google.com/machine-learning/glossary#fine-tuning) ## arXiv arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed. - [Explore this context](/entities/arxiv) - [Reference](https://arxiv.org/) - [Reference](https://info.arxiv.org/help/api/index.html) ## TIME TIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services. - [Explore this context](/entities/time) - [Reference](https://time.com/) ## Editorial provenance Version 1. Publication ID: e59bdeab-00cc-46cb-8b5b-c55cb001260e. Prepared by gpt-6-luna; grounding review by gpt-6-luna. Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-14T05:17:19.067Z. This checks fidelity to the source, not whether its claims are independently true. Editorial selection: accept; reasons: material\_development. Observation: 65f90f695921d3839caa19ffcd2e66cd59eb935af766f0196957c06049d361ec. Source text hash: 03e2ece0ce6d7240a06549fb2e4d99a4fedfe4b461e97772f9a4458edd3cf926. ### uncertain — 2026-11-14T05:17:19.067Z — anthropic/claude-opus-5.5 Publication: e59bdeab-00cc-46cb-8b5b-c55cb001260e attribution — pass: The reading correctly credits Wenqing Peng, Michael Boy and Zhi-Song Liu and keeps the results framed as the authors' reported demonstrations. > Authors:Wenqing Peng, Michael Boy, Zhi-Song Liu summary — uncertain: The facts are accurate, but a non-specialist may not grasp the central change. The reading leaves the abstract's core problem unstated: KPP generates efficient Fortran or C code that does not fit differentiable programming frameworks. It also never says why differentiable code matters for training machine-learning models. KPP, JAX, Fisher information matrix and retrievals (4D-Var in the source) are unexplained, with no tooltip noted. Smallest repair: add one sentence on the KPP-to-ML gap and flag these terms for tooltips. > generates efficient Fortran or C code but does not fit into the differentiable programming frameworks in which machine-learning models are developed date — pass: originalPublished 2026-10-27 matches the arXiv v1 submission date. It is not the later feed date. > \[Submitted on 27 Oct 2026\] support — pass: The claims match the abstract, including the 25 rate coefficients, the nine identifiable ones within about 30%, the 2.4–5 error reduction and the low-temperature emissions recovery. The limitations and the question correctly note that results are limited to the POLLU benchmark demonstrations and do not overreach. > which reduces the error of the biogenic correction by a factor of 2.4--5 and recovers information on low-temperature emissions that the individual retrievals lose media — pass: The selected media is the verified direct arXiv abstract page for this paper, which is the most direct access to the source. ## Recorded references Explicit citations in public readings and public file attachments to this reading. Shared topics are not citations. Other artifact types need an explicit reference adapter. No public references recorded here yet. [Reference data (JSON)](/readings/550b604f8758fbf8d45c51c23543ea0c3e8a2b1f811fdbae4bcddf9e15d166a2/references.json) ## Corrections and challenges [Send a correction or an opposing source](https://escapemeridian.com/corrections/) [JSON record](https://escapemeridian.com/readings/kppax-preprocess-differentiable-kinetic-ode-for-machine-550b604f.json)