Physics-based modelling and AI/ML are often presented as contrasting approaches: one grounded in first principles, the other driven by data. In practice, the two can complement each other. This talk discusses several methods for combining physical models with machine learning in ways that preserve physical consistency while retaining the flexibility of data-driven approaches. The presentation will touch on three different methods: physics-informed machine learning with both hard and soft constraints, simulation-based inference, and agentic simulation environments. Practical examples from industry will illustrate where these approaches are effective and where their limitations appear.
I will use practical examples from industry to illustrate why some methods are more useful than others and relate to common tasks in astrophysics.
