Simulation you can trust
Stiff dynamics: when a simulation quietly breaks
A cytokine that is consumed within minutes and T cells that change over days force a choice of time step. With the wrong step, a standard solver matched the reference for two weeks and then diverged. An implicit solver stayed stable with steps 70 times larger, at some cost in accuracy
Dose-response fitting: why more points can make a fit worse
A polynomial through 13 evenly spaced doses predicted responses from -40% to +140%. Adding doses made it worse. Better dose spacing or a monotone curve fixed it.
Surrogate models: fast, but only where trained
A small neural network learned a 30-day Treg simulation from 150 runs. Inside the trained region its error is small; outside it is roughly 35 times larger. Every surrogate should come with its validity region.
Capacity planning: where is the bottleneck?
For an illustrative assay core, linear optimization found the weekly mix that maximizes samples analysed (489) and showed that thermocycler time, plate-reader time and reagent budget limit the output, while technician time does not.
All models are illustrative and not fitted to client or patient data.
