Inferring dissipation maps from videos using convolutional neural networks

Quantifying entropy production (EP) is essential to understand stochastic systems at mesoscopic scales, such as living organisms or biological assemblies. However, without tracking the relevant variables, it challenging figure out where and what extent EP occurs from recorded time-series image data experiments. Here, applying a convolutional neural network (CNN), powerful tool for processing, we develop an estimation method through unsupervised learning algorithm that calculates only movies. Together with attention map of CNN's last layer, our can not quantify but also produce spatiotemporal pattern (dissipation map). We show accurately measures creates dissipation in two nonequilibrium systems, bead-spring model elastic filaments. further confirm high performance even noisy, low spatial resolution data, partially observed situations. Our will provide practical way obtain maps ultimately contribute uncovering nature complex systems.