🤖 machine learning

SpinCastML an Open Decision-Making Application for Inverse Design of Electrospinning Manufacturing: A Machine Learning, Optimal Sampling and Inverse Monte Carlo Approach

SpinCastML is an open-source, chemically informed machine learning and Inverse Monte Carlo framework that enables the inverse design of electrospinning processes by predicting full fiber diameter distributions and generating feasible polymer-solvent parameters to achieve user-defined nanofiber outcomes.

Original authors: Elisa Roldan, Tasneem Sabir

Published 2026-08-14

📖 3 min read☕ Coffee break read

CC BY 4.0

Original authors: Elisa Roldan, Tasneem Sabir

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a world where you could snap your fingers and conjure a tiny, invisible net made of fibers so thin they are thinner than a human hair. This isn't magic; it's a manufacturing technique called electrospinning. Think of it like a high-tech, electric version of a cotton candy machine. Instead of spinning sugar, scientists shoot a stream of liquid plastic through a needle. Then, they zap it with a powerful electric charge. This charge stretches the liquid into a super-thin thread that whips around wildly before hardening into a fiber. These fibers can be woven into mats used for everything from bandages that help wounds heal to filters that clean the air we breathe.

But here's the tricky part: making these fibers is a bit like trying to bake the perfect cake without a recipe. You have to mix the right ingredients (different types of plastics and liquids), set the oven temperature (the electric voltage), and control how fast you pour the batter (the flow rate). If you change just one tiny thing—like adding a drop more liquid or turning up the voltage slightly—the whole result can change. You might end up with perfect, smooth threads, or you might get a messy blob of beads. For decades, figuring out the perfect settings has been a game of "guess and check," where scientists spend weeks mixing and testing just to get the fiber thickness they need. This is slow, wasteful, and frustrating.

Enter SpinCastML, a new digital tool designed to stop the guessing game. The researchers behind this tool built a massive "library" of knowledge by collecting data from nearly 1,800 different scientific studies, gathering over 68,000 individual measurements of fiber sizes. They didn't just look at the average size; they looked at the whole picture, including how much the sizes varied. Using this library, they taught a computer brain (a machine learning model) to understand the complex dance between the ingredients and the electric settings. But the real magic is how they use it: instead of asking the computer, "What will happen if I do this?", they ask the reverse: "I want fibers this specific size; what settings should I use?"

The paper introduces SpinCastML, an open, free software application that acts as a reverse-engineering guide for making these tiny fibers. The researchers found that by using a clever sampling method to balance their data and a specific type of computer model called Cubist, they could predict fiber sizes with high accuracy (getting a score of over 0.92 out of 1.0). More importantly, they built a "probabilistic" engine, which means it doesn't just give one single answer. Instead, it understands that there are many different ways to make the same fiber size. It generates a list of safe, chemically possible recipes, telling the user not just what to mix, but how likely it is to work.

In a real-world test, the team used SpinCastML to design a recipe for a specific type of plastic called PVA. The software suggested a mix of water and plastic, along with specific voltage and flow settings, to create fibers around 250 nanometers wide. When the researchers actually went into the lab and tried it, the fibers they made were almost exactly what the computer predicted, landing right in the target range. This proves that the tool can successfully turn a desired outcome (a specific fiber size) into a practical, working recipe, potentially saving scientists months of trial-and-error and helping them design better materials for medicine, energy, and the environment.

Technical Summary: SpinCastML – An Open Decision-Making Application for Inverse Design of Electrospinning Manufacturing

Problem Statement
Electrospinning is a versatile manufacturing technique for producing micro- and nanoscale fibrous materials, yet its rational design remains hindered by the process's intrinsic multivariate and nonlinear nature. Small variations in polymer chemistry, solvent selection, or operating conditions can shift electrohydrodynamic regimes, leading to distinct fiber-diameter distributions. Crucially, functional performance often depends on the shape and variability of these distributions rather than the mean diameter alone. Current design practices rely heavily on time- and material-intensive trial-and-error optimization. Furthermore, existing machine learning (ML) approaches often predict single scalar outcomes (mean diameter) rather than full distributions, and few tools address the "inverse design" problem: determining the specific formulation and processing parameters required to achieve a desired fiber morphology. Additionally, the lack of chemical feasibility constraints in design loops can lead to proposed solutions that are experimentally impossible (e.g., insoluble polymer-solvent pairs).

Methodology
The authors introduce SpinCastML, an open, standalone executable application that implements a distribution-aware and chemically constrained inverse-design framework. The methodology integrates four key components:

Data Curation and Preprocessing: A rigorously curated dataset comprising 68,480 individual fiber-diameter measurements extracted from 1,778 experimental studies across 16 polymers was assembled. Data were harmonized to standardize units, solvent nomenclature, and concentration notations.
Optimal Sampling Strategies: To address high-dimensional heterogeneity and material imbalance (where certain polymers are overrepresented), the framework compares three sampling strategies: simple random sampling, Sobol+D-optimal sampling, and polymer-balanced Sobol+D-optimal sampling. The latter combines space-filling low-discrepancy sequences (Sobol) with statistically informative subset selection (D-optimal) while enforcing stratification to ensure minority polymers are adequately represented without losing internal variability.
Machine Learning Benchmarking: Eleven ML algorithms (including Cubist, GBM, XGBTree, Random Forest, SVM, and linear models) were systematically benchmarked using cross-validation (k=3, 5, 10) and independent test sets. Performance was evaluated using RMSE, MAE, MAPE, and $R^2$ . The study emphasizes selecting models that generalize well across chemically diverse systems rather than overfitting to specific polymers.
Inverse Monte Carlo (IMC) Engine: Unlike deterministic optimization, SpinCastML formulates inverse design as a probabilistic inference problem. The IMC engine generates chemically feasible configurations (polymer-solvent-process) that satisfy compatibility constraints (solubility and miscibility) while assigning quantified success probabilities. It operates in two modes:
- Experimental Mode: Samples from the empirical distribution of historical data to reproduce known behaviors.
- Optimization Mode: Generates synthetic configurations within observed ranges to explore the design space, filtering candidates based on strict chemical feasibility rules (e.g., polymer-solvent solubility ratings: OK, COND, NO).

Key Contributions

Distribution-Aware Modeling: SpinCastML explicitly models complete fiber-diameter distributions rather than single-point estimates, enabling the assessment of process robustness and the probability of meeting user-defined targets.
Chemically Constrained Inverse Design: The framework embeds chemical feasibility (polymer-solvent solubility and solvent-solvent miscibility) directly into the design loop, ensuring that generated solutions are experimentally actionable.
Open, Reproducible Tool: The application is released as a standalone executable (Shiny-based) with a curated dataset, allowing researchers without ML expertise to upload data, train models, and perform inverse simulations.
Hybrid Sampling Framework: The development and validation of a polymer-balanced Sobol+D-optimal sampling strategy that effectively mitigates data imbalance while preserving high-dimensional coverage.

Results

Model Performance: The Cubist model, trained under the polymer-balanced Sobol+D-optimal sampling strategy, achieved the most robust global predictive performance ( $R^2 > 0.92$ , RMSE = 260 nm) with stable generalization across diverse polymers. Linear models (e.g., LM, glmnet) performed poorly ( $R^2 \approx 0.40$ ), confirming the necessity of nonlinear learners for electrospinning.
Sampling Impact: The balanced sampling strategy significantly reduced discrepancies between cross-validation and test metrics compared to random or unbalanced Sobol+D-optimal sampling, indicating reduced overfitting and improved generalization.
Inverse Simulation: The IMC engine successfully reconstructed experimental distributions with $R^2 > 0.90$ . In the "Optimization" mode, the system could identify families of feasible conditions. For some polymers (e.g., Polystyrene), optimization sharpened predictions and increased success probabilities (up to 99.9%), while for others (e.g., PVA), it revealed the breadth of feasible conditions or highlighted limitations in finding narrow targets.
Prospective Case Study: A case study targeting Poly(vinyl alcohol) (PVA) fibers with a diameter of 250 nm ( $\pm$ 50 nm) demonstrated the workflow's efficacy. The IMC identified 20 chemically feasible candidate families. One condition (12% PVA in water, 25 kV, 0.3 mL/h) was selected for experimental validation. The resulting fibers had a mean diameter of 243.59 nm, falling within the predicted target window, and the experimental distribution closely matched the IMC-predicted distribution.

Significance and Claims
The paper claims that SpinCastML advances electrospinning from empirical parameter tuning toward quantitatively guided, application-driven design. By reframing electrospinning as a distributional inverse problem and embedding chemical feasibility directly into the design loop, the work establishes a new standard for how electrospun processes are designed and validated under realistic manufacturing variability.

The authors emphasize that the primary contribution is not merely incremental improvement in prediction accuracy, but a conceptual shift: moving away from deterministic "single best recipe" optimization toward probabilistic, decision-ready guidance that accounts for experimental uncertainty and chemical constraints. This approach is presented as particularly relevant for emerging challenges such as blended polymer systems, green solvent selection, and application-specific nanofiber architectures where simple forward prediction is insufficient. The tool is designed to lower the technical barrier to advanced modeling, fostering reproducible and auditable research in the field.

More like this