基于深度学习的微纳米纤维膜制备工艺逆向预测研究OA

Micro-nano fibrous network membranes exhibit highly tunable functional properties that are intrinsically linked to their microstructural characteristics,which are governed by complex,nonlinear relationships with processing parameters during fabrication.Conventional parameter optimization relies heavily on iterative trial-and-error experiments,a process that is both time-consuming and resource-intensive,particularly when dealing with multiple coupled variables such as spinning pressure,nozzle diameter,and feeding rate.To transcend these limitations,this work introduces a deep learning-enabled inverse design strategy that directly infers optimal process parameters from scanning electron microscopy SEM)images of target microstructures.By establishing a data-driven mapping between structural features and process conditions,the proposed approach substantially reduces experimental iterations and associated costs. A representative system using polyethylene oxide PEO)was employed to validate the proposed framework.A series of microfluidic spinning experiments were conducted under systematically varied conditions,producing seven distinct types of fibrous membranes F1-F7,each characterized by a unique combination of process parameters.High-resolution SEM imaging was performed for each sample type,yielding over 700 images per category.To enhance the model's robustness and generalization,an extensive image augmentation pipeline was implemented,including multi-axis rotation,center cropping,random erasing with probability p=0.5,and color perturbation brightness adjusted by 0.2,contrast by 0.1 to simulate operational variances and imaging artifacts. An inverse prediction model was constructed based on a refined ResNet-50 architecture.Transfer learning was incorporated using weights pre-trained on ImageNet1K V2 to alleviate overfitting and improve feature extraction capability especially critical given the relatively small dataset.The model was further enhanced by integrating a dynamic channel-wise attention mechanism that amplifies the influence of salient microstructural attributes such as fiber intersections,pore boundaries,and network homogeneity.The original classifier was replaced with a fully connected layer tailored to the seven process categories.The network was trained using the AdamW optimizer under a cosine annealing learning rate schedule with label smoothing applied to the cross-entropy loss function to improve generalization. Interpretability analyses using Grad-CAM revealed that the model's attention shifted across network depths:shallow layers focused on elementary features like fiber edges and pores,middle layers on junction connectivity and local density while deeper layers synthesized higher-order information related to diameter distribution and network orientation.This work validates the feasibility of the proposed inverse design approach.When tested on a previously unseen process setting F8 the model accurately predicted three out of four parameters,with only a minor error in spinning pressure 0.04 MPa underscoring its strong extrapolation capability. Following training,the proposed model,achieved a training accuracy of 99.04%and a validation accuracy of 96.35%,with an F1-score consistently maintained at 0.9 640.Misclassifications,as visualized through a confusion matrix,were primarily confined to adjacent process categories e.g.,between F1 and F2,indicating that the model successfully discerned subtle structural differences corresponding to parameter variations.To verify practical applicability the model was deployed to predict parameters for validation samples.Independent t-tests confirmed that no significant difference existed between the fiber diameters of original and predicted samples 15.95±0.88 μm vs.16.03±0.61 μm t 41)=-0.480,p=0.634).Bland-Altman analysis further indicated a mean bias of only 0.08 μm with 95%of data points within the limits of agreement. This study presents a robust,image-driven inverse design methodology that effectively bridges microstructural patterns and process parameters for micro-nano fibrous materials.The combination of deep learning,attention mechanisms,and advanced data augmentation offers a viable and efficient to conventional empirical optimization,with significant implications for accelerating the development and intelligent manufacturing of functional fibrous assemblies.The framework is generalizable and can be adapted to other material systems where process-structure-property relationships play a decisive role.