Machine-Learning-Based Dispersion Optimizer for Carbon Nanotubes across Dispersant–Solvent–Process Space

Appropriately pairing dispersants
and solvents and tuning processing
conditions yield high-quality, stable carbon nanotube (CNT) dispersions;
however, experimental investigation of this combinatorial design space
is slow and nonsystematic. We report a machine-learning-based CNT
dispersion optimizer that simultaneously considers dispersant, solvent,
and the dispersion process to predict CNT dispersibility and crystallinity.
The data set comprised 666 dispersions (36 organic dispersants, 22
solvents, and two dispersion methods) with systematic variations in
composition and processing parameters. Dscore and IG/ID quantified the dispersion quality and structural integrity, respectively.
Using molecular descriptors, experimental variables, and solvent–dispersant
similarity metrics as inputs, an eXtreme Gradient Boosting (XGBoost)
model achieved a coefficient of determination (R2) = 0.57, mean absolute error (MAE) = 0.08 for Dscore and R2 = 0.73, MAE =
9.84 for IG/ID. These correspond to mean absolute errors below 10% of the target
ranges, indicating sufficient performance for screening-grade formulation
design. Limited quantitative accuracy was displayed for dispersants
outside the training set; solvent-dependent trends were reproduced,
and practically useful formulations were identified. Virtual screening
within the learned domain yielded improved formulations. SHapley Additive
exPlanations and feature-group ablation revealed that Dscore was governed primarily by solvent–dispersant
compatibility encoded by similarity and distance-like descriptors,
while IG/ID was dominated by process intensity. These elements constitute a
CNT dispersion optimizer that efficiently prescreens formulation and
processing conditions and can be extended to other nanomaterial dispersion
systems. This prescreening framework reduces empirical trial-and-error
and promotes solvent- and process-constrained formulation design by
treating dispersions as enabling intermediate materials for the downstream
manufacturing of films, fibers, and composites.

合理搭配分散剂与溶剂并调控加工条件，可制备得到高质量、稳定的碳纳米管（CNT）分散液；然而，针对该组合设计空间开展实验研究既耗时又缺乏系统性。本研究提出一种基于机器学习的碳纳米管分散液优化模型，可同时兼顾分散剂、溶剂与分散工艺，用以预测碳纳米管的分散性与结晶度。该数据集包含666组分散液样本，涵盖36种有机分散剂、22种溶剂以及2种分散方法，其组成与加工参数均经过系统性变量设置。其中Dscore与IG/ID分别用于量化分散液的质量与结构完整性。以分子描述符、实验变量以及溶剂-分散剂相似性指标作为输入，极限梯度提升（eXtreme Gradient Boosting, XGBoost）模型在Dscore预测任务中取得了决定系数（R²）=0.57、平均绝对误差（MAE）=0.08的结果；在IG/ID预测任务中则实现了R²=0.73、MAE=9.84的性能。上述误差均低于目标区间的10%，表明该模型足以满足筛选级配方设计的性能需求。对于训练集之外的分散剂，模型的定量精度有所局限；但模型可复现溶剂依赖的变化趋势，并能筛选出具备实用价值的配方。在模型习得的参数空间内开展虚拟筛选，可获得性能更优的配方。SHapley可加性解释（SHapley Additive exPlanations, SHAP）与特征组消融实验表明，Dscore主要受基于相似性与类距离描述符所表征的溶剂-分散剂相容性调控，而IG/ID则主要由加工强度决定。该模型可作为碳纳米管分散液优化工具，高效完成配方与加工条件的预筛选，且可推广至其他纳米材料的分散体系研究中。该预筛选框架可减少经验性试错流程，通过将分散液视为薄膜、纤维与复合材料下游制造的功能性中间材料，推动兼顾溶剂与工艺约束的配方设计工作。