# Analysis of Multiscale Condensation Phenomena Using a Zero‐Shot Computer Vision Framework

## TL;DR

A new computer vision method automatically analyzes droplet behavior during condensation without needing labeled data, offering insights into energy and water systems.

## Contribution

A zero-shot computer vision framework using SAM for label-free segmentation of condensation dynamics is introduced.

## Key findings

- The framework detected over one million droplets with high accuracy without annotated datasets or retraining.
- Machine learning predicted condensation rates with a mean absolute percentage error of 7.8%.
- Surface properties like contact angle hysteresis were shown to govern droplet behavior and population statistics.

## Abstract

Understanding and controlling condensation is central to diverse energy and water systems, yet quantifying its inherently multiscale droplet dynamics has remained elusive. Here, a label‐free computer vision framework that leverages the Segment Anything Model (SAM) for zero‐shot segmentation is demonstrated, achieving high accuracy by systematically detecting more than one million droplets without any annotated datasets or retraining. The framework extracted statistical features such as droplet radius, number of coalescences, mean coalescing diameter, growth rate, and condensation mass from microscale to macroscale observations. This enabled direct visualization of dynamic condensation cycle including nucleation, growth, coalescence, sweeping, and renucleation, while also revealing the role of surface properties, where contact angle hysteresis governed droplet departure size, morphology, and population statistics in agreement with classical models. To extend beyond characterization, a machine learning model was trained with Pearson correlation–selected parameters to predict condensation rates under diverse environmental conditions, achieving a mean absolute percentage error of 7.8%. These frameworks highlight the potential of artificial intelligence to understand dynamic phase‐change mechanisms, and to guide the design of advanced surfaces and systems for thermal management, desalination, and water harvesting.

A zero‐shot computer vision framework quantifies multiscale condensation dynamics by automatically segmenting droplets and extracting physical parameters without labeled data. The workflow integrates data mining and statistical analysis to reveal droplet growth, coalescence statistics, and sweeping behaviors, enabling label‐free measurement of heat flux and condensation mass through artificial intelligence–assisted analysis.