Two AI Methods Battle It Out to Predict Smog — One Wins

Researchers compared two methods for pinning down reaction rate coefficients in a toy-case autoxidation mechanism, a challenge for explicit atmospheric chemical mechanisms where many reaction pathways are only indirectly observed through high-resolution mass spectrometry. The test used synthetic data with known ground truth: an ODE-constrained neural-network optimiser, which gives fast point estimates of uncertain rate coefficients, and Markov Chain Monte Carlo, which samples the posterior distribution and quantifies parameter uncertainty. With unperturbed and low-noise synthetic observations, both converged on the known rate coefficients, and the neural-network optimiser was faster. But under high-noise conditions, at a signal-to-noise ratio of roughly S/N = 1, MCMC was substantially more robust at recovering them. The posterior analysis showed that mass-spectral aggregation broadens credible intervals even at low noise, and that high-noise mass spectra can leave many individual reaction rates weakly identifiable. Still, posterior predictive validation showed that broad parameter uncertainty constrained by MCMC remained consistent with accurately reproducing the observable mass spectrum. The takeaway: the two are complementary — neural-network optimisation works well on informative data, while MCMC is essential for diagnosing uncertainty, non-uniqueness, and identifiability in noisy or aggregated inverse problems.