Prediction of deep learning algorithms for the microemulsion generation conditions in complex environments

Background

Constructing high-quality datasets for AI-driven microemulsion prediction remains challenging due to limitations in chip design, imaging hardware, and fluid dynamics. Conventional semi-empirical models suffer from poor accuracy and generalizability in complex nonlinear systems, while traditional microfluidic chips often yield non-spherical droplets. Optical constraints in capillary-based systems further hinder data acquisition.

Methods

A deep learning framework integrates three components: (1) A multi-branch CNN with residual modules and self-attention (RAM-CNN) for robust droplet/fluid morphology recognition; (2) GAN-LGBMnet, combining adversarial networks and LightGBM, to augment small datasets and analyze key features; (3) A multi-output neural network (MO-DNN) predicting single/double-emulsion parameters. The system is deployed via an intuitive GUI (MICA) for non-specialist use.

Significant Findings

RAM-CNN achieves 100 % fluid-regime and 95.8 % droplet-morphology recognition accuracy, maintaining >91 % performance under ±28 % brightness variations. Enhanced by GAN-LGBMnet, MO-DNN predicts droplet diameters and generation rates with MAPE <7 % for both emulsion types. The MICA platform demonstrates <10 % error on unseen data, enabling precise emulsion design. This work bridges theoretical models with practical microfluidic optimization through automated, user-friendly AI tools.