You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Stiff DDE support: MethodOfSteps with implicit solvers needs improvement #3150
Very stiff DDE systems (e.g., host-parasite dynamics with large parameter values) require extremely low tolerances to produce correct results, making solves prohibitively slow. The current implicit solvers (Rosenbrock, Rodas, TRBDF2) wrapped in MethodOfSteps lack the specialized stiff DDE handling that dedicated solvers like RADAR5 provide.
Key areas for improvement:
Better discontinuity handling for stiff problems
More robust step size control near discontinuities
The Robertson DDE benchmark (from Guglielmi's paper) demonstrated instabilities with Rosenbrock methods at longer integration intervals, and issues with history function extrapolation near discontinuities.
MWE (stiff host-parasite DDE)
using DelayDiffEq, OrdinaryDiffEq
# Host-parasite model where S > ~500 requires extremely low tolerances# Full MWE at: https://github.com/SciML/DelayDiffEq.jl/issues/198# With Rodas5 and careful dtmax/tolerance tuning, solutions can be found,# but the parameter space exploration is fragile.
A dedicated stiff DDE algorithm (e.g., based on RADAR5's approach with residual control and smoothness indicators) would significantly improve robustness.
I'd like to take this. A RADAR5 wrapper is out: its license (manual §9) is research and teaching only and not redistributable, so no JLL, and a port has to come from the papers.
DelayDiffEq already tracks breaking points (tstops, and a root search on rejected steps). RADAR5 goes further: it solves for the breaking point together with the Radau stages, also searches when the error jumps, and includes the dense-output error in its step control.
I'd first benchmark RADAR5 against MethodOfSteps(RadauIIA5()) on #198 and the long-interval #67 Robertson to see which of those matters. @ChrisRackauckas sound right?
Activity
singhharsh1708 commentedon Nov 18, 2026
I'd like to take this. A RADAR5 wrapper is out: its license (manual §9) is research and teaching only and not redistributable, so no JLL, and a port has to come from the papers.
DelayDiffEq already tracks breaking points (tstops, and a root search on rejected steps). RADAR5 goes further: it solves for the breaking point together with the Radau stages, also searches when the error jumps, and includes the dense-output error in its step control.
I'd first benchmark RADAR5 against
MethodOfSteps(RadauIIA5())on #198 and the long-interval #67 Robertson to see which of those matters. @ChrisRackauckas sound right?