Figure 8.

Learned bivariate posterior for the system state escape rates ${\lambda }_{esc}$ and FRET efficiencies ${ϵ}_{FRET}$ for synthetic data with crosstalk. The ground truth is shown with red dots. In (a), we show the learned posterior using the model that does not correct for crosstalk consistently shows deviation from the ground truth with higher FRET efficiency estimates on account of more donor photons being detected as acceptor photons. The FRET efficiencies estimated by our sampler for this case are ${0.314}_{-0.009}^{+0.007}$ and ${0.135}_{-0.005}^{+0.002}$, and the predicted escape rates are ${1.90}_{-0.11}^{+0.17}$ ms⁻¹ and ${1.05}_{-0.18}^{+0.10}$ ms⁻¹. For the corrected case shown in (b), FRET efficiencies estimated by our sampler are ${0.276}_{-0.010}^{+0.006}$ and ${0.088}_{-0.006}^{+0.004}$. Furthermore, predicted escape rates are ${1.85}_{-0.14}^{+0.15}$ ms⁻¹ and ${1.06}_{-0.10}^{+0.12}$ ms⁻¹. The corrected model mitigates this bias as demonstrated by the posterior in (b).