When lots of data is available, but there's little prior knowledge about model parameters, convergence of ABC algorithms is typically enhanced through gradually decreasing the tolerance between simulated and measured data, akin to simulated annealing (SA) in optimization. SA-ABC (short SABC) employs an adaptive annealing schedule that attempts to minimize entropy production, which is a measure for wasted computation, within the particle ensemble. SABC has shown superior performance in many applications. It requires little tuning and does not suffer from an N^2 overhead, unlike other ABC algorithms that gradually decrease the tolerance, which is a big advantage on parallel computing infrastructures. After an introduction to SABC, I will show new inference results, for stochastic solar dynamo models calibrated to radionuclide time series. I will also report on our experience with machine learning techniques, for the generation of summary statistics from time-series.