Transfer learning for emulation of hydrodynamic simulations

Extracting the properties of quark-gluon plasma from data taken at the Large Hadron and Relativistic Heavy Ion Colliders requires computationally expensive simulations and Bayesian statistical methods. Bayesian inference requires millions of model simulations which is computationally prohibitive. One, therefore, constructs surrogates that can be trained on data from a sparse set of full model simulations. Obtaining the training data from the full simulations then dominates the computational cost of the analysis. When the model is improved, recalibrating its parameters in principle entails the same cost again. We introduce a computationally much less expensive alternative to build accurate emulators for the improved model with a much smaller number of additional training data, by transferring knowledge about the model parameters from emulators trained on the large set of existing training data for the original model. We validate the new method by building emulators for simulations for different collision systems and collision energies and for different viscous corrections at particlization. We show that the method is surprisingly efficient. Limitations and further improvements of the technique will be discussed.