Fig. 6: Automated mapping of multi-dimensional phase diagrams.

A Two independent experiments (Supplementary Figs. 46–47) were conducted to explore the effect of salt (NaCl) on the phase behavior of poly-L-(lysine)₁₀₀ and poly-L-(aspartic acid)₂₀₀. The combined dataset, made from 1760 datapoints, was used to construct the “ground truth” three-dimensional phase diagram, here reported. Iso-probability surfaces indicate phase separation (blue, higher opacity) and no phase separation (red, lower opacity). B Four distinct orientations of the phase diagram with non-transparent surfaces are shown to emphasize phase behavior from different perspectives. C Balanced accuracy plot showing the accuracy on the prediction for each successive cycle with respect to the “ground truth” phase diagram in panel A. Cycle 0 represents the balanced accuracy computed with respect of a randomly generated phase diagram as a baseline comparison. D Within-experiment Jensen-Shannon Divergence (JSD) plotted across cycles. This metric tracks convergence by comparing consecutive cycles, illustrating how each replicate approaches the final phase diagram. Cycle 0 reflects divergence from a randomly generated phase diagram. E Between-experiment Jensen-Shannon Divergence (JSD) across replicates at each cycle. Similar to panel D, Cycle 0 serves as a baseline, representing divergence from a randomly generated phase diagram. Total polypeptide consumption for 1280 samples: 85.1 mg poly-L-(aspartic acid) and 82.1 mg poly-L-(lysine).