Uncertainty-Guided Framework: Surrogate Model Enhancement and Inverse Design for Microfluidic Droplet Generators

Microfluidic droplet generation is widely used in drug delivery, materials synthesis, and biochemical analysis, but efficiently designing devices in high-dimensional parameter spaces remains a computational challenge. The complexity of multiphase flow dynamics makes high-fidelity simulation costly, and inverse design relying solely on simulation is practically infeasible. A recent study proposes a unified approach that couples adaptive surrogate model enhancement with simulation-driven inverse design, offering a new way out of this dilemma.

Surrogate Models and Uncertainty Quantification

At the core of the study is the use of neural network ensembles to predict flow regime and droplet diameter. High-fidelity Volume-of-Fluid (VOF) simulations provide ground-truth data, while the ensemble quantifies epistemic uncertainty through consensus level (for flow regime classification) and coefficient of variation (for diameter prediction).

A key finding is that uncertainty concentrates near flow regime transition boundaries, where physical behavior is highly sensitive to parameter perturbations. This is intuitive—when the system is in a critical state between two flow regimes, small parameter changes can lead to completely different outcomes. The ensemble's predictions diverge most in these regions, precisely revealing the physically sensitive zones.

This uncertainty quantification is not only a diagnostic tool but also serves as a compass guiding data collection. Based on this, the researchers propose a physics-informed enrichment strategy that deploys additional targeted simulations in these critical regions and retrains the surrogate models in the ensemble, while simultaneously mitigating class imbalance.

The Volume-of-Fluid (VOF) method is a mainstream interface-capturing technique in numerical simulation of multiphase flows. It implicitly describes the position of phase interfaces by defining the volume fraction of each phase within every cell on a fixed grid (e.g., liquid fraction α, ranging from 0 to 1), without explicitly tracking interface geometry. A cell with 0<α<1 is an interface cell. The advantage of VOF lies in its ability to naturally handle topological changes of droplets (such as breakup and coalescence) and its good mass conservation, making it particularly suitable for simulating droplet generation, stretching, and breakup in microfluidics. The cost is that interface reconstruction requires fine grids, and the time step is constrained by capillary stability conditions; a single high-fidelity simulation often takes hours to days, which is the fundamental reason this article emphasizes that "high-fidelity simulation is costly" and "inverse design relying solely on simulation is infeasible."

Epistemic uncertainty and aleatoric uncertainty are two types of uncertainty from different sources in machine learning. The former arises from insufficient model knowledge and can theoretically be reduced by adding training data; the latter arises from inherent noise or randomness in the data itself and cannot be eliminated by adding data. In this article, the consensus level and coefficient of variation measured by the neural network ensemble quantify epistemic uncertainty—the greater the divergence among ensemble members for the same input, the less the model "knows" about that region. This distinction is crucial because only epistemic uncertainty can be reduced by targeted simulation. The enrichment strategy is effective precisely because it targets the portion of uncertainty that can be reduced through additional data, rather than the inherent random fluctuations of the physical system.

Performance Gains from the Enrichment Strategy

Validation on an independent test dataset shows significant improvement of the enriched ensemble model. The most intuitive metric: the proportion of droplet diameter predictions with errors in the 0-3% range nearly doubles, while high-error cases (over 10%) drop substantially.

The significance of this result is that it demonstrates that uncertainty-guided targeted simulation is more efficient than uniformly increasing simulation data. Computational resources are concentrated in the most physically sensitive regions where the model is most uncertain, achieving higher prediction reliability at lower cost. For microfluidic design, this means engineers can significantly shorten design cycles without sacrificing accuracy.

Class imbalance is a common problem in classification tasks: when the number of samples across classes differs greatly, the model tends to favor the majority class, degrading minority-class recognition performance. In microfluidic flow regime classification, different flow regimes (such as squeezing, dripping, and jetting) occupy regions of varying sizes in the parameter space, and some flow regimes naturally have few samples. If simulation data are generated by uniform sampling, the sample size for rare flow regimes may be far from sufficient for the ensemble model to learn adequately, and their classification boundaries will be blurrier.

The physics-informed enrichment strategy, when adding simulations to high-uncertainty regions, further exacerbates this imbalance—because samples near transition boundaries tend to concentrate in a few flow regimes. Therefore, researchers need to adopt mitigation measures simultaneously, such as weighting minority-class samples, resampling, or introducing class-balance terms in the loss function. This step is often overlooked, but without it, the overall accuracy improvement of the classifier from enrichment may come at the expense of rare flow regime recognition rates, thereby affecting downstream inverse design judgments about feasible flow regimes. Understanding this helps grasp why the enrichment strategy must also emphasize "mitigating class imbalance" rather than simply adding simulation data.

Particle Swarm Optimization-Driven Inverse Design

Building on the enhanced surrogate model, the study proposes an inverse design framework integrating particle swarm optimization (PSO). The framework searches for physically realizable configurations that meet target droplet sizes under constraints, while balancing prediction accuracy, prediction confidence, and physical consistency.

The difficulty of inverse design is that a target droplet size may correspond to multiple parameter combinations, but not all are physically feasible or stable in actual fabrication. By incorporating the surrogate model's prediction confidence into the optimization objective, the framework can avoid high-risk regions where the model is uncertain, favoring configurations that are reliably predicted and physically consistent. Validation shows that the inverse design workflow can reliably identify physically realizable configurations meeting target droplet sizes.

Particle swarm optimization (PSO) is a population-based, gradient-free global optimization algorithm proposed by Kennedy and Eberhart in 1995, inspired by the foraging behavior of bird flocks. The algorithm maintains a set of "particles," each representing a candidate solution in the parameter space with a position and velocity. In each iteration, particles update their velocity and position based on their own historical best position and the swarm's global best position, thereby converging toward promising regions of the search space. PSO does not require gradient information of the objective function, is simple to implement with few parameters, and is suitable for the inverse design problem in this article where the objective function is given by the surrogate model and may be non-convex and constrained.

When using PSO for inverse design, the objective function is not a single droplet diameter error but a composite objective integrating prediction accuracy, prediction confidence, and physical consistency. Specifically, the predictive mean given by the surrogate ensemble serves as the accuracy term, while the ensemble variance (or coefficient of variation) serves as a confidence penalty term, causing the optimization process to actively move away from high-risk regions where the model is uncertain. This design makes PSO tend to converge to parameter combinations that both meet the target size and lie within the surrogate model's reliable prediction range, thereby alleviating the "solution exists but is unreliable" problem in inverse design. The limitation of PSO is that it may converge prematurely in high-dimensional constrained spaces; in practice, hyperparameters such as inertia weight and learning factors often need tuning, or combination with local search.

Methodological Significance and Applicability Boundaries

The value of this work is not limited to microfluidic droplet generation. It demonstrates a general paradigm: use uncertainty quantification to identify weak spots in the surrogate model, use targeted simulation to enrich training data, and then embed the enhanced surrogate model into an optimization loop to complete inverse design. This paradigm can theoretically be transferred to a broader range of microfluidic system optimization problems.

Of course, the effectiveness of the method depends on the coverage of initial simulation data and the calibration quality of the ensemble model. Uncertainty estimates themselves may be biased; if the correlation between epistemic uncertainty and true error is weak, the benefits of the enrichment strategy will be discounted. In addition, the convergence of PSO in high-dimensional constrained spaces also needs to be adjusted for specific problems.

From an engineering perspective, the truly valuable aspect of this research is that it turns "where to spend computational budget" into a principled decision problem. The uncertainty-guided enrichment strategy gives a clear basis for allocating simulation budget, rather than relying on researchers' intuition or uniform sampling. For teams with limited computational resources, this approach has direct practical reference value.

One-line Summary

The study enhances surrogate models with uncertainty-guided simulation enrichment and combines particle swarm optimization to achieve reliable inverse design of microfluidic droplet generators.

Paragraph Summary

Addressing the difficulty of designing microfluidic droplet generators in high-dimensional parameter spaces, this study proposes a method that unified couples adaptive surrogate models with simulation-driven inverse design. The core is to use neural network ensembles to predict flow regime and droplet diameter, and to quantify epistemic uncertainty via consensus level and coefficient of variation, finding that uncertainty concentrates at flow regime transition boundaries. Based on this, a physics-informed enrichment strategy is proposed to deploy targeted VOF simulations in high-uncertainty regions and retrain, while mitigating class imbalance. After enrichment, the proportion of droplet diameter predictions within 0-3% error nearly doubles, and high-error cases drop substantially. On this basis, the study embeds the enhanced surrogate model into a particle swarm optimization inverse design framework to find configurations that meet target droplet sizes, are physically realizable, and reliably predicted under constraints. This paradigm can be transferred to broader microfluidic optimization, but its effectiveness depends on initial data coverage and uncertainty calibration quality.