KPPax puts a neural network inside an atmospheric chemistry calculation, learning emission corrections from simulated concentrations. The 27 October preprint translates reaction mechanisms into differentiable JAX code.[1]
Atmospheric reactions consume some gases and create others, at very different speeds. A concentration records the effect of that interacting system. The original KPP paper shows how a written reaction scheme becomes executable equations, including sunlight-driven changes and ozone formation and destruction. The older software already supported calculations tracing how changes to inputs affect outputs.[2]
The experiment represents human and natural emission sources. It first estimates corrections separately for individual scenarios, then trains a network on those answers. Joint fine-tuning instead adjusts the network through the chemistry across many scenarios.[1]
With noisy observations of five gases, biogenic log-emission-correction root-mean-square error fell from 0.309 to 0.128 within training ranges. This evaluates corrections against known truth in a 20-species box model with known initial conditions and no transport.[1]
An emission and a later concentration are different quantities. A gas may be produced or consumed through reactions before its concentration is recorded. Following those pathways makes the inference problem intelligible: the calculation must recover what entered the system from what remains after chemistry has acted.[2]
Sources & further reading
- KPPax: preprocess differentiable kinetic ODE for machine learning in atmospheric chemistry arXiv · 27 Oct 2026 · Accessed 29 Oct 2026
- Technical note: Simulating chemical systems in Fortran90 and Matlab with the Kinetic PreProcessor KPP-2.1 Atmospheric Chemistry and Physics · 26 Jan 2006 · Accessed 29 Oct 2026