Inference pipeline using amortized neural posterior estimation. The joint simulation model p(x, θ) = p(θ)p(x|θ) is used to generate a training set {(θ, x)} of model parameters θ and exoplanet spectra observations x. A conditional normalizing flow p_ϕ(θ|x) composed of an embedding network and three invertible transformations t_i is trained to estimate the posterior density p(θ|x). Once trained, sampling from the posterior estimator is as fast as a forward pass through the normalizing flow. Inference can be repeated for many observations without having to regenerate data nor retrain the normalizing flow.