Accelerated Alkyd Emulsification Through AI-Driven Experimentation

By Ioanna Tzortzi, Albin Jansfelt, and Marie Westerblad, Perstorp AB; Nils Honhon, Université Libre de Bruxelles; and Eirini Palaiologlou, National Technical University of Athens

Introduction

Waterborne alkyd binders remain an important resin technology in the coatings industry, offering a viable route to low-VOC formulations with high renewable-content potential while preserving the application characteristics and aesthetic properties traditionally associated with solventborne systems.¹ As fatty-acid-modified polyesters, alkyds can reach very high renewable carbon content depending on raw material selection. However, their conversion into stable oil-in-water (O/W) emulsions remains a significant thermodynamic and kinetic challenge. Establishing and maintaining colloidal stability under storage and thermal stress requires precise control of interfacial phenomena and is still commonly approached through extensive empirical screening of emulsifier chemistry and process conditions.

This challenge is amplified by the strongly path-dependent nature of alkyd emulsification. Emulsion formation and subsequent emulsion stability do not depend on composition alone, but on the coupled interaction between alkyd chemistry, surfactant structure, degree of neutralization, temperature, and hydrodynamic processing trajectory.² The resulting formulation domain is therefore highly nonlinear, and difficult to navigate by conventional trial-and-error experimentation. Moreover, no universal compatibility rules exist: surfactant systems that perform well for one alkyd family may fail for another because of differences in polarity, molecular architecture, and rheological behavior. Waterborne alkyd development must therefore be treated as a constrained multi-objective formulation problem in which successful emulsification, particle-size control, and resistance to accelerated aging must be satisfied simultaneously before downstream coating performance can even be assessed.

Machine learning (ML) and materials informatics offer a promising route to accelerate formulation development,³ yet their application to complex heterophase and macromolecular systems remains comparatively limited relative to real-world formulation challenges. In coatings, recent studies have demonstrated the utility of active or sequential learning in narrower and more compositionally fixed systems, including self-healing epoxy coatings and automated lacquer-formulation workflows.^4,5 These studies are important demonstrations of ML-guided optimization, but they primarily address local optimization within a single coating platform and a relatively bounded formulation space. In contrast, the present work addresses a different level of formulation complexity: rather than optimizing one fixed resin system, it aims to learn transferable formulation–process relationships across multiple chemically distinct alkyds within a large, sparse, and dynamically expanding design space during the ML-guided optimization. The challenge here is therefore not only efficient candidate selection, but predictive generalization across heterogeneous resin environments under industrially and application relevant formulation and process constraints.

In this work, we report on a closed-loop active-learning workflow for waterborne alkyd emulsification and, to our knowledge, the first such implementation in this specific coatings context. We constructed a dataset that transforms formulation and process knowledge into quantitative, model-ingestible descriptors, including alkyd physicochemical fingerprints, rheology-derived variables, process-history features, and molecular descriptors computed from surfactant SMILES (Simplified Molecular Input Line Entry System) strings. This framework moves beyond static quantitative structure–property relationship approaches by implementing iterative train–propose–validate–retrain cycles designed to balance exploration of under-sampled regions with exploitation of high-probability candidates. Importantly, the variables considered here were not selected as an idealized benchmark system, but were defined from experimentally relevant formulation and process dimensions encountered in real alkyd-emulsification development.

A central novelty of the present study lies in the scale and dynamic nature of the formulation problem addressed. The initial campaign, spanning 24 alkyds, already corresponded to a formulation–process space on the order of 50,000 candidate combinations of alkyd identity, emulsifier system, degree of neutralization, and temperature levels. After 12 experimental cycles, expansion to 48 alkyds increased this accessible space to on the order of 100,000 combinations. Yet only a minor fraction of this domain was experimentally sampled, progressing from 160 executed experiments over ∼50,000 (≈0.32% coverage) theoretical combinations to 500 experiments over ∼100,000 combinations (≈0.50% coverage). Thus, rather than solving a fixed optimization problem, the model was required to learn within a moving and progressively broader search space while retaining practical candidate-ranking value under extreme sparsity. In this sense, the contribution of the present work is not merely the application of ML to an emulsion system, but the demonstration that a chemistry-aware closed-loop workflow can generate high-quality candidates without requiring dense experimental coverage of a large, heterogeneous, and dynamically expanding formulation domain.

To evaluate practical formulation success, we adopted a rigorous four-index criterion comprising: (1) successful emulsification, (2) fresh particle-size specification, (3) stability after accelerated thermal aging, and (4) particle-size retention after aging. Within this framework, the present study positions waterborne alkyd emulsification as a coatings-formulation informatics problem at the interface of colloidal engineering and data-driven materials design, while establishing a methodology for active-learning-guided development under realistic formulation complexity and sparse experimental coverage.

 

Continue reading in the July-August 2026 Issue of CoatingsTech.