Artificial intelligence application for rapid fabrication of size-tunable PLGA microparticles in microfluidics

The study addresses the challenge of generating uniform, size-controlled polymeric microparticles, a key factor for stability and performance in applications such as biomimicry, drug delivery, and pharmaceutics. Traditional prediction of droplet and resultant particle sizes in microfluidics using non-dimensional parameters requires prior knowledge of fluid properties (e.g., viscosity, surface tension), which is difficult to ascertain, making in silico prediction using AI attractive. Artificial intelligence, particularly machine learning with artificial neural networks (ANNs), is well-suited to model nonlinear relationships. The context includes advances in biomimetic membrane models and polymeric vesicles, with PLGA being a widely used, FDA-approved polymer for drug delivery. Prior ML efforts in microfluidics have focused on single systems or specific parameters (e.g., nanoprecipitation, droplet stability). The objective here is to develop ANN models capable of predicting droplet and particle sizes for PLGA across three distinct microfluidic systems (single- and multiple-emulsion formats), individually and in combined models, to enable rapid, accurate, and generalizable in silico size prediction.

The paper situates its work within AI/ML applications across domains including pharmaceutics and microfluidics. It notes previous ML models targeting microfluidic-related predictions such as nanoprecipitation of drug particles, droplet stability, and fluid/flow parameters. It also reviews the utility of microfluidics for generating monodisperse particles with tunable properties, and the advantages of polymeric vesicles and PLGA (biocompatibility, biodegradability, FDA approvals) for biomedical applications. The literature underscores the complexity and nonlinearity of relationships governing microfluidic droplet formation and particle solidification, supporting the choice of ANNs for predictive modeling.

Chemicals and materials: Glass microfluidic flow-focusing chips (Dolomite Microfluidics), PLGA (75:25, Mw 76,000–115,000), dichloromethane (DCM), PW11 surfactant blend, Aqua-Phase (1–5% proprietary polymer blend). Flow control via Mitos P-Pumps and in-line flow rate sensors; high-speed microscopy for in-junction imaging. Experimental systems: Three microfluidic systems (MFS A, B, C) were built and optimized.

• MFS A (single-junction, hydrophilic): 3D flow focusing device, channels 100 µm deep, 105 µm wide at junction; three inlets, one outlet. Continuous phase (aqueous with 2% PW11) injected in side inlets; dispersed phase (PLGA in DCM) injected centrally. Droplets formed at the junction, collected in aqueous phase, then solidified by DCM evaporation to form particles. Flow ranges explored: continuous phase 10–150 µL/min; dispersed phase 0.5–31.5 µL/min; outside these, co-laminar/single-phase flow occurred. PLGA concentrations tested: 1–20%.
• MFS B (seven parallel junctions, hydrophilic): Telos 3D flow focusing device with 7 parallel junctions; junction depth 54 µm, width 75 µm. PLGA concentrations 0.5–7.5%. Continuous and dispersed phase flow rates were roughly 70× higher than in MFS A.
• MFS C (two chips in series for W/O/W multiple emulsions): First chip hydrophobic (14×17 µm junction) forms water-in-PLGA/DCM droplets (0.5 wt% Synperonic PE/F68 stabilizer). Output routed via 0.25 mm ID FEP tubing to second hydrophilic chip (as in MFS A) to form W/O/W droplets with outer aqueous (Aqua-Phase). Final particles formed via staged DCM evaporation (faster at droplet edges). Resulting particles washed and imaged. Imaging and measurements: Droplet sizes tracked at junction by high-speed imaging; particle sizes measured off-chip. Example: in MFS A, droplets 95.23 ± 1.79 µm shrank to 31.29 ± 1.08 µm post-evaporation. In silico models: Multilayer perceptron (MLP) ANNs implemented in Statistica v13.3. Five models: ANN-A (MFS A), ANN-B (MFS B), ANN-C (MFS C), ANN-AB (combined A+B), ANN-ABC (combined A+B+C). For each model, 1000 MLPs were trained; best networks selected by highest r² and lowest sum-of-squares error. Inputs: PLGA concentration, continuous-phase flow rate (AP FR), dispersed-phase flow rate (PLGA FR), and type (droplet or particle). Combined models also included chip type. Continuous inputs normalized to 0–1. Data split per model into training (~60%), test (~20%), validation (~20%); external datasets used for independent testing. Dataset sizes: ANN-A (n=186; external n=20), ANN-B (n=25; external n=3), ANN-C (n=12; external n=2), ANN-AB (n=211; external n=23), ANN-ABC (n=223; external n=25). Activation functions evaluated: identity, logistic, tanh, exponential; hidden layer size optimized by trial-and-error. Outputs: measured droplet/particle sizes. Sensitivity analyses performed to assess relative importance of inputs.

• Microfluidic generation achieved highly monodisperse PLGA droplets and particles in single and multiple emulsions.
• MFS A (single junction): Increasing PLGA concentration (1–20%) increased droplet and particle sizes. At 1% PLGA, droplets 35–101 µm yielded particles 12–23 µm; at 2%: droplets 26–88 µm, particles 16–28 µm; at 10%: droplets 44–86 µm, particles 28–40 µm; at 20%: droplets 46–48 µm, particles 35–44 µm. Sizes decreased with increasing aqueous flow rate and decreasing PLGA (dispersed) flow rate; the converse produced larger sizes. Flow ranges used: continuous 10–150 µL/min; dispersed 0.5–31.5 µL/min.
• MFS B (7 junctions): Droplets 26–102.5 µm and particles 7–40 µm across PLGA 0.5–7%. Flow rate variations showed expected trends similar to MFS A.
• MFS C (two-stage W/O/W): Inner water droplets 3–14 µm; outer PLGA shells 57–72 µm; final particles after DCM evaporation 30–45 µm. Inner droplet number and size controlled by flow conditions: higher PLGA flow in first chip increased inner droplets per PLGA shell; higher outer aqueous flow in second chip increased final droplet size.
• ANN model performances (from Table 1 and figures): • ANN-A (5-7-1; hidden tanh, output identity): r² training/test/validation = 0.988/0.988/0.986; errors 3.29/2.57/3.80 µm. External r² = 0.990. Residuals typically ±5 µm. • ANN-B (5-4-1; hidden exponential, output identity): r² = 0.998/0.990/0.994; errors 0.54/5.23/4.94 µm. External r² = 0.944. Residuals typically ±5 µm. • ANN-C (5-10-1; hidden exponential, output tanh): r² = 0.999/1/1; errors 0.48/0.88/0.83 µm. Residuals typically ±2 µm. High r² attributed to small dataset (n=12), with potential overfitting; nevertheless, good agreement with external data. • ANN-AB (7-12-1; hidden exponential, output exponential): r² = 0.969/0.945/0.948; errors 9.05/13.49/17.09 µm. External r² = 0.972. • ANN-ABC (8-9-1; hidden exponential, output exponential): r² = 0.978/0.975/0.971; errors 6.23/6.45/9.46 µm. External r² = 0.953.
• Example shrinkage upon solvent evaporation (MFS A): droplet size 95.23 ± 1.79 µm reduced to 31.29 ± 1.08 µm.
• Sensitivity analyses: • ANN-A: Importance order: aqueous phase flow rate > PLGA concentration > PLGA flow rate > droplet/particle type. • ANN-B: Aqueous phase flow rate > PLGA flow rate > PLGA concentration > droplet/particle type. • ANN-C: PLGA concentration > aqueous phase flow rate > PLGA flow rate > droplet/particle type. • ANN-AB: Chip type > aqueous phase flow rate > PLGA flow rate > PLGA concentration > droplet/particle type. • ANN-ABC: Chip type (dominant; sensitivity > 5000) > PLGA flow rate ≈ aqueous phase flow rate > PLGA concentration > droplet/particle type. The broader ANN-ABC revealed substantial importance of dispersed-phase flow rate for determining sizes.
• Overall, simple-structure ANNs yielded accurate, rapid predictions of droplet and particle sizes across distinct microfluidic systems, including combined-system models.

The findings demonstrate that ANNs effectively capture the complex, nonlinear relationships between PLGA concentration, flow rates of continuous and dispersed phases, chip characteristics, and the resulting droplet/particle sizes in microfluidic generation. By building separate models for each device (ANN-A, ANN-B, ANN-C) and then progressively integrating datasets (ANN-AB, ANN-ABC), the study shows strong generalization across different microfluidic architectures (single- and multiple-emulsion formats) and operating regimes. High correlations between predicted and observed sizes in training, test, validation, and external datasets support the models’ predictive utility. Sensitivity analyses elucidate how key factors influence size outcomes, with aqueous-phase flow rate being dominant in single-chip systems and chip type becoming most influential in multi-system models, while dispersed-phase flow rate emerges as particularly important in the diverse ANN-ABC model. These results address the research aim by enabling rapid in silico prediction of size-tunable PLGA microparticles, reducing empirical optimization and facilitating design and scale-up across different microfluidic platforms.

This work presents an AI/microfluidics application that enables rapid, accurate prediction of size-tunable PLGA microparticles across multiple microfluidic systems. Five ANN models were developed, culminating in a general ANN-ABC model that integrates data from three distinct devices to predict droplet and particle sizes with high accuracy. The models provide mechanistic insight via sensitivity analyses, highlighting the roles of flow rates, polymer concentration, and chip type. This approach can streamline experimental optimization, saving time and costs, and is potentially extendable to other polymers, droplet contents, and device designs. Future research could expand datasets to further enhance generalization, extend to additional materials and complex emulsions, and integrate real-time control for closed-loop microfluidic fabrication systems.

The dataset for ANN-C (multiple emulsions) was small (n=12), which may have led to overfitting despite good agreement with external data. Additionally, some external validation sets were limited in size (e.g., ANN-B external n=3; ANN-C external n=2), constraining assessment of broader generalizability. For MFS A, droplet formation did not occur outside specified flow ranges (continuous 10–150 µL/min; dispersed 0.5–31.5 µL/min), limiting predictive applicability beyond these operational windows.