Combining mechanistic models with ML

Mechanistic models encode biological processes explicitly, while machine-learning models can learn patterns or unknown relationships from data. Hybrid models combine these strengths.

Ways to combine them

ML can estimate parameters or unobserved quantities used by a mechanistic model, approximate computationally expensive simulations, learn a model discrepancy, or replace an uncertain component while known mechanisms remain equation-based.

Example

In an epidemic model, the compartment equations may remain mechanistic while a data-driven component estimates a time-varying transmission rate from mobility, behaviour or environmental measurements.

Careful validation

A hybrid model can still overfit and can inherit misspecification from both components. Its predictive and mechanistic claims should therefore be evaluated separately.