PITCHFORK: Emulating individual mode frequencies of solar-like oscillators using a neural network

Characterising stellar fundamental parameters is difficult. For instance, generating new stellar evolutionary tracks to reflect observations is computationally expensive. Additionally, quantifying the uncertainty inherent in comparison to stellar model grids can be challenging. To mitigate this issue, we use a neural network as an emulator, converting discrete grids of stellar models into continuous functions with easily quantifiable emulation uncertainties. Once trained, this emulator is capable of rapidly predicting a host of observable quantities including individual mode frequencies, each with easily quantifiable emulation uncertainties. Trained emulators are then used for likelihood estimation in Bayesian inference to recover plausible estimates of the posterior distributions. We present results for inferred fundamental parameters of solar- like oscillators. We observe that the recovered posteriors for simulated stars may exhibit non-Gaussian distributions, multimodality, and strong covariance. Furthermore, we present results obtained from real stars. Finally, we show how covariance in emulation prediction residuals should be accounted for, and discuss the improvements from using ensemble approaches to emulation in the future.