Rapid Model-Based Recipe Design from Limited-Sample Datasets: Experimental Validation in Nanoparticle and Microgel Systems

Abstract

In data-constrained experimental domains such as nanoparticle engineering, microgel synthesis, and pharmaceutical formulation, researchers frequently face the challenge of modeling systems governed by complex and highly nonlinear relationships among variables. These applications often involve limited datasets due to the cost, time, and resource demands of generating new samples, making conventional trial-and-error approaches inefficient. As a result, there is a growing need for data-driven methodologies that can reliably predict product behavior and guide recipe design using minimal experimental input. This thesis presents a series of strategies to enhance prediction accuracy and reliability in small datasets through localized modeling, quantitative model reliability assessment, and guided expansion of the available dataset. The first contribution involves coupling Latent Variable Modeling (LVM) with clustering to create local Partial Least Squares (PLS) models tailored to subsets of similar samples. This combination simplifies the underlying data structure by reducing multicollinearity via projection into latent space and grouping structurally similar data, thereby improving prediction fidelity. The framework was validated using the prediction of the Volume Phase Transition Temperature (VPTT) of dual-responsive microgels—a property influenced by several formulation variables—with results that showed significantly improved prediction accuracy. Building on these advances, the second contribution focuses on the inverse problem of design space identification—determining input configurations that are most likely to yield desired output properties. To do this robustly, the Prediction Reliability Enhancing Parameter (PREP) is introduced, a novel metric that unifies multiple LVM alignment diagnostics including Hotelling T², Squared Prediction Error (SPE), and score alignment factors into a single predictive reliability score. PREP is calibrated in a data-driven, case-specific manner and facilitates the ranking of candidate formulations by their expected predictive reliability. Extensive validation on simulated datasets has demonstrated that PREP significantly accelerates the identification of optimal solutions, particularly under highly nonlinear conditions and limited data regimes. PREP was subsequently deployed across real experimental case studies involving the formulation of nanoparticles and microgels. In one study, a microgel with a tightly constrained particle size of ~100 nm was successfully designed from an initial dataset spanning sizes of 170–900 nm, with PREP delivering a near-target solution in minimal iterations while competing design approaches failed under the applied constraints. In another case, PREP enabled the identification of polyelectrolyte complexes with particle sizes below 200 nm and polydispersity indices under 0.2, again demonstrating superior efficiency and accuracy relative to conventional approaches. Overall, this work offers a practical and scalable pathway for predictive modeling and recipe design in settings constrained by data scarcity and high experimental costs. The methodologies developed—particularly the integration of local LVM models and the PREP-based design space identification—can be broadly applied to other high-value domains requiring precision formulation and optimization such as drug delivery, nanomedicine, and advanced materials development.