PhaseXplorer Creates High-Dimensional Phase Diagrams with Closed-Loop Active Learning

Phase separation fundamentally governs material properties and cellular function on multiple organizational scales. Conventional approaches to studying this nonlinear phenomenon necessitate resource-intensive experiments. As such, investigations were limited to low-dimensional space. We present PhaseXplorer, a platform that combines microfluidics, microscopy, and machine learning to efficiently study phase separation systems. PhaseXplorer autonomously designs, generates, and analyzes samples in a closed-loop active learning workflow until an accurate phase diagram is obtained. Using an acquisition function that balances exploration and exploitation, all the phase boundaries are located with minimal sampling. A convolutional neural network executes real-time image recognition to identify microfluidic droplets and phase separation within them in less than 1 ms per droplet. PhaseXplorer standardizes analysis across experiments and does not require calibration nor extensive postexperiment analysis. We demonstrate PhaseXplorer's capabilities using a poly rA model system by creating a four-dimensional phase diagram 100 times faster than traditional methods while simultaneously consuming 10,000 times less material.