The energy release efficiency and service safety of energetic materials are closely related to their crystal morphology and microstructure. However, an inherent contradiction has long existed between their high energy density and low mechanical sensitivity. In recent years, the preparation of spherulites and other aggregated morphologies by antisolvent crystallization has become an important technical pathway to resolve this contradiction. Compared with traditional acicular crystals, the unique radial branching structure of spherulites can significantly reduce impact sensitivity and friction sensitivity, while effectively improving powder flowability, enhancing packing density, and increasing thermal stability. Therefore, they show outstanding application potential in the crystallization engineering of energetic materials.

Nevertheless, achieving precise regulation of spherulite morphology still faces significant challenges. The spherulite growth process depends on a non-classical branching mechanism and is strongly influenced by the multiscale nonlinear coupling of factors such as supersaturation evolution trajectory, fluid shear, and interfacial adsorption of additives. Current regulation strategies mostly rely on empirical trial and error, which is inefficient in high-dimensional process parameter spaces. At the same time, due to the high risk of energetic materials, relevant experimental data are extremely scarce. Therefore, how to break through multivariable coupling constraints under small-sample conditions and achieve a methodological leap from experience-driven approaches to predictive inverse design has become a core scientific problem that must be solved in this field.

To address the above issues, a joint team composed of the Institute of Chemical Materials, China Academy of Engineering Physics, Chongqing University, and others conducted in-depth research using ZEPTOOLS' ZEM series台式 scanning electron microscope, and constructed a machine learning and constrained inverse design framework for small-sample data. They quantitatively revealed the nonlinear regulation mechanism of key process parameters on morphology and successfully achieved the precise directional construction and process optimization of LLM-105 energetic material from acicular to dense spherulites.

Title:From acicular to Spherulite: crystallization optimization guided by machine learning for enhanced energetic material properties 
Journal:Chemical Engineering Science 
URL:http://dx.doi.org/10.1016/j.ces.2026.124959

During the antisolvent crystallization of energetic materials, multiphase transport and crystal growth kinetics are highly intertwined, making crystal morphology difficult to predict, and the scarcity of experimental samples further restricts the application of traditional modeling methods. To achieve accurate discrimination of microscopic morphology under small-sample data constraints, the study took the typical energetic compound LLM-105 (2,6-diamino-3,5-dinitropyrazine-1-oxide) as the object and used polysuccinimide (PSI) as a crystal morphology modifier. It systematically obtained 107 batches of antisolvent crystallization experimental data, among which 83 expert-annotated samples were rigorously divided into three categories: spherulites, acicular crystals, and other non-ideal phases. Using five key process parameters, including solution concentration, solvent-to-antisolvent volume ratio, additive ratio, antisolvent temperature, and stirring rate, as physical input features, multiple machine learning classification models were compared and evaluated. Tests showed that the CatBoost classification algorithm achieved an accuracy of 0.8235, a weighted F1 score of 0.816, and an area under the receiver operating characteristic curve of 0.941 on the test set, demonstrating excellent generalization ability and stability in small-sample scenarios and successfully establishing a high-precision mathematical mapping from process parameters to crystallization morphology categories.

To break the long-standing black-box effect in the regulation of spherulite growth and clarify the hierarchical regulation mechanisms of various process variables, the study used the game-theoretic SHAP method to deeply analyze the internal decision logic of the machine learning model. Model interpretability analysis clearly revealed that solution concentration and solvent/antisolvent volume ratio constitute the dominant thermodynamic driving force regulating morphology evolution, jointly contributing about 65% of the model decision variance. Among them, high concentration and low solvent/antisolvent ratio can synergistically provide extremely high local instantaneous supersaturation, thereby inducing non-classical branching growth. The polysuccinimide additive dosage and antisolvent temperature, as secondary regulatory factors, accounted for about 30% of the decision variance. The modifier molecules finely regulate the compactness and microscopic branching characteristics of the aggregates by selectively competitively adsorbing on specific crystal faces and pinning growth steps, while the stirring rate had a relatively weak effect within the investigated range. This analysis successfully transformed scattered qualitative experience into a quantitative decision-making basis for guiding process regulation.

On the basis of clarifying the causal relationships of key processes, the study further proposed an exhaustive screening inverse design method incorporating physical feasibility constraints to replace traditional blind trial and error. By using the trained classifier as a surrogate model, it completed a comprehensive traversal of more than 640,000 feasible discrete state spaces after filtering out non-physical conditions and directly locked onto the process combination with the highest probability of forming spherulites. In the subsequent microstructure verification of samples, the research team used ZEPTOOLS' ZEM series台式 scanning electron microscope at an accelerating voltage of 10 kV to systematically observe the crystal surface morphology of the prepared samples. This directly confirmed that the optimal predicted process could successfully suppress the coarse acicular growth of the raw material and instead form typical spherical micro-nano aggregates with clear branching radiating from the center and a compact structure, fully verifying the reliability and executability of the discrete-constraint-based inverse screening technology in the directional reconstruction of energetic material crystal morphology.

The successful reconstruction of microscopic aggregate morphology directly promoted significant improvement in the macroscopic properties of the powder. The research team further analyzed the structure-activity relationship between spherulite configuration evolution and powder functional properties. Phase and pore testing showed that, while maintaining the original monoclinic crystal form unchanged, the successfully constructed dense spherulites increased the specific surface area from 2.18 m²/g for the raw material to a maximum of 9.44 m²/g. The pore distribution at the nano and mesoscale was highly consistent with the density of microscopic branching, clarifying the essential law by which the input of micromixing energy per unit volume regulates microscopic growth density. This work not only provides a precise process path for the efficient spherulitization of LLM-105, but its overall framework based on conventional process variables, quantitative microscopic features, and lightweight machine learning models also provides a general and transferable model-assisted design paradigm for crystal morphology control of more energetic materials and specialty fine chemicals facing data scarcity challenges.

The ZEPTOOLS ZEM series台式 scanning electron microscope is a highly integrated, portable, and cost-effective scientific research instrument. It features rapid vacuuming, high imaging speed, and a variety of signal detector options, making it suitable for morphology observation and composition analysis, and it can also adapt to various in-situ experimental needs. The equipment has low installation environment requirements, does not require special floor selection, and is simple to operate. Non-specialists can quickly get started with it, allowing it to be widely applied in many fields such as new material research and development, life sciences, failure analysis, and industrial quality inspection, providing a powerful microscopic characterization solution with both high performance and high cost-effectiveness for scientific research institutes and enterprise users.