Right answer for the wrong reasons: a cautionary tale from mix experimental modeling

Mix of high-Z material into the fuel is a significant degradation mechanism for fusion experiments, yet is often observed via indirect means. For inertial fusion, the ‘separated reactant’ technique is a popular experimental platform. We fit a radiation-hydrodynamics model including a turbulent mix model to previous data, obtaining an excellent match to observables, to predict a new experimental series. Unfortunately this model missed every experimental point by at least a factor of three due to missing physics in the model, clearly indicating that the training data was insufficient and the pre-shot model got the right answer for the wrong reasons. Adding a second mechanism (diffusion) to the model plus additional data gives a statistically significant discrimination between the two mechanisms [A.B. Zylstra et al., Phys. Rev. E 97, 061201(R) (2018)]. This is a cautionary tale about the underlying bias often present in models from our preconceptions, which will be a challenge for applying ML methods to physical science.