Sensitivity Analysis
Sensitivity analysis measures how a model’s outputs respond to changes in its inputs or parameters. Local methods use derivatives at a nominal point (computed by finite differences, direct or adjoint sensitivity equations, or automatic differentiation), while global methods such as Sobol variance decomposition assess importance over a whole uncertain parameter space; hybrid methods such as DELSA evaluate local sensitivities at many samples to map how importance varies across that space.
Local and hybrid sensitivity of reaction networks with AD (2022)
Yang, Achar and Kitchin treated the degree of rate control as a local sensitivity and computed it with automatic differentiation, avoiding the step-size trade-off of finite differences and the extra assembly needed with direct or adjoint ODE sensitivities (Yang et al. 2022). They then used the distributed evaluation of local sensitivity analysis (DELSA), with 200 quasi-random Sobol samples of 17 uncertain rate constants, to obtain histograms of each parameter’s importance to the rate of propylene oxide formation. Only three steps had mean importance above 0.1, matching the nominal DRC analysis, but one of them was important in about half the samples and unimportant in the other half, information a single local analysis cannot provide. Summary: yang2022b.